Dylan Di Filippo
Dylan Di FilippoAI journal

Journal IA 2026-W41

Understand what changes. Choose what deserves a test.

← 2026-W402026-W42 →Browse all weeks →

Understand what changes.

Public sources. Verified facts. Analysis kept separate from announcements.

24 stories

AnthropicModels

Claude Haiku 5.5 targets fast tasks and sub-agents

Anthropic launches a new small model oriented high volumes, latency and cost, with adaptive reasoning and adjustable effort.

available

Facts, analysis and limits

Verified facts

  • Model identifier: claude-haiku-5-5.
  • Context of 1 million tokens and maximum output of 128,000 tokens according to API notes.
  • Pricing depends on the total prompt size: up to 100,000 tokens, input is $0.10 and output $0.50 per million tokens; above that threshold, input is $0.50 and output $2.50 per million. This is not incremental pricing applied only to excess tokens.
  • Available via Claude Platform and main cloud; Claude Code supports it in v2.1.293.
  • Migration: old thinking/sampling/prefill parameters can cause 400 errors.

What changes

A candidate for large-scale abstracts, classifications and subtasks.

Practical implication

Test accuracy and cost per task and follow the migration guide before replacement.

Availability

Claude Platform, AWS, Google Cloud, Azure; cloud regions and contractual access to check for Belgium/EEA.

Limits and uncertainty

Official launch of 7 October, no email reception of 8. Opus/Sonnet 5.5 remain launches of September. Performances and savings announced by Anthropic.

Introducing Claude Haiku 5.5 \ Anthropic ↗Claude Platform release notes - Claude Platform Docs ↗
Read full article →
OpenAIModels

GPT-6 provides an interactive interface to ChatGPT Chat

Answers can combine text, visuals, graphics and small interactive tools, and begin before the end of the reasoning.

global rollout

Facts, analysis and limits

Verified facts

  • Global deployment from October 7 for Plus, Pro, Business and Enterprise; Free and Go from October 8.
  • Sol for the paid plans listed, Luna for Free/Go.
  • Intelligent UI from Instant to Extra High; Pro reasoning remains on Astra without this interface.
  • Work and Codex models do not change with this announcement.

What changes

The answer becomes an object to explore, not just a paragraph to read; checking the data behind a graph remains indispensable.

Practical implication

Try a simple comparison or calculator and check the values with an external source.

Availability

Global deployment, therefore Belgium/EEA included in the announced scope; customers and Enterprise parameters applicable.

Limits and uncertainty

These are Chat versions of October, distinct from the September Sol/Luna used in Work/Codex; not a new GPT-6.1 Sol launch.

ChatGPT release notes | OpenAI Help Center ↗GPT-6 Sol and GPT-6 Luna: October 2026 update - OpenAI Deployment Safety Hub ↗
Read full article →
AnthropicApplications

Claude launches Dashboards and Motion; Docs, Slides and Design come out of beta

Dashboards creates connected dashboards and Motion editable animations in code; other tools become accessible including on Free.

beta and general availability

Facts, analysis and limits

Verified facts

  • Dashboards in beta on paid plans; Motion on Team and Enterprise.
  • Export MP4 for Motion, which is not based on a video generation model.
  • Enterprise: Dashboards and Motion disabled by default; Docs/Slides/Design activated by default on October 15.
  • The design standalone site is scheduled to close on December 14th; migration of the design systems proposed.
  • Each dashboard chart shows its query and last refresh; named connections include Redshift, BigQuery, ClickHouse, Databricks, Snowflake and Salesforce.
  • The standalone Design site's closure on December 14 also ends its public links; its chats and comments will no longer be available.

What changes

Claude brings together documents, data visualization and animation, but the useful point is the possibility of inspecting the numbers and revising the content.

Practical implication

Open the query behind an important number and check its update; save useful Design comments and chats before December.

Availability

Plans as listed; Belgium/EEA scope is not explicitly detailed and administrator policies apply.

Limits and uncertainty

Do not confuse Claude native tools with Google Workspace add-ons; Design has existed in conversations since September. Dashboards are presented as exploratory and complementary to the BI, not as a guaranteed replacement of validated reports.

Build live dashboards and animate explainers with Claude | Claude by Anthropic ↗Build live dashboards and animate explainers with Claude | Claude by Anthropic ↗Build live dashboards and animate explainers with Claude | Claude by Anthropic ↗
Read full article →
GoogleVoice & creativity

Nano Banana 2.1: generation and conversational retouch

The new version of the image model is available in the Gemini API.

General availability

Facts, analysis and limits

Verified facts

  • Identifier is gemini-nano-banana-2.1.
  • Google announces better visual quality, better monitoring of instructions, character consistency and text rendering.
  • The 1:4, 4:1, 1:8 and 8:1 ratios are documented in 1K, 2K and 4K.
  • gemini-3.1-flash-image is deprecated, without announced shutdown date.

What changes

Conversational editing and panoramic formats broaden visual uses; a beautiful demonstration does not prove fidelity to your content.

Practical implication

Compare the same brief and edit across multiple trials, including text and character consistency.

Availability

Gemini API / AI Studio: Belgium is one of the supported regions. Effective access, billing and model-specific restrictions remain to be checked. This list does not establish a European processing residence and does not apply to all functions of the Gemini application.

Limits and uncertainty

Quality improvements presented by Google; no independent comparison made here.

Release notes  |  Gemini API  |  Google AI for Developers ↗Gemini API : régions disponibles ↗
Read full article →
MistralModels

Mistral Large 4: preview API, weights announced for late October

Mistral's multimodal model targets code, agents and professional uses; weights are not yet delivered to the source.

Public preview

Facts, analysis and limits

Verified facts

  • ML4 is described as a one-trillion-parameter model with 52 billion active parameters.
  • A public preview API is available in Mistral Studio.
  • Weights are promised by the end of the month, during a red teaming phase.
  • Mistral states that training and the preview service use its own European infrastructure.

What changes

A European offer seeks to combine capacity, controlled infrastructure and future distribution of weights; the preview is not yet an autonomous deployment.

Practical implication

Judge the preview on useful use and wait for weights and license before concluding on actual autonomy.

Availability

Preview served on European infrastructure according to Mistral; no Belgian exclusion mentioned. Check region, contract and account availability.

Limits and uncertainty

Do not write "weight available": the October 6th post announces them for the end of the month. Benchmarks not reverified independently.

Introducing Mistral Large 4 | Mistral ↗Introducing Mistral Large 4 | Mistral ↗
Read full article →
AnthropicApplications

Claude works in Google Docs, Sheets and Slides

An add-on adds Claude to Google files and connectors allow you to create and edit them from Claude.

public beta

Claude for Google Workspace™. Official demonstration, not an independent test. Press play to start.
Open official video ↗Demonstration source ↗
Facts, analysis and limits

Verified facts

  • All Claude paid plans.
  • Default mode Ask before edits; mode Accept all edits optional.
  • Existing Google permissions apply.
  • Team/Enterprise: connectors to be activated by an owner.
  • Sheets: formulas, pivot tables, native charts and range cleanup through Python; Slides: reuse of layouts and checks for overflowing elements.
  • Enterprise controls, including Compliance API, CMEK and OpenTelemetry export audit, also apply to add-on.

What changes

The choice of the assistant separates itself from the choice of the office suite: teams can keep their Google documents without just adopting Google's AI.

Practical implication

Try a non-sensitive shared document revision in Ask before edits mode and compare the quality to the tools already used.

Availability

All paid plans; no exclusions from Belgium/EEA indicated in the announcement, actual availability depending on deployment and Google/admin rules.

Limits and uncertainty

Official date: October 6th, not the 7th that corresponds to the newsletter relay. Allowing a connector does not exempt you from reviewing file sharing or controlling proposed changes.

Claude now works with Google Docs, Sheets, and Slides | Claude by Anthropic ↗Claude now works with Google Docs, Sheets, and Slides | Claude by Anthropic ↗
Read full article →
OpenAISafety & policy

OpenAI prepares text watermark in the EU

textGrain adds an invisible statistical signal to text; the API offers opt-in, while ChatGPT/Codex are to introduce it gradually in the European Union.

announced deployment and API opt-in

Facts, analysis and limits

Verified facts

  • Opt-in global API for some models, disabled by default.
  • Deployment of ChatGPT and Codex in the EU over the following weeks.
  • Access to the detector initially reserved for approved researchers and expert bodies.

What changes

A concrete change in outputs for European users.

Practical implication

Avoid any disciplinary use based solely on a detector result.

Availability

Belgium affected by the future EU deployment; do not automatically extend to all EEAs.

Limits and uncertainty

The watermark does not prove the accuracy or the author; retouching and translation can weaken detection.

Our approach to EU text provenance rules | OpenAI ↗
Read full article →
AnthropicApplications

Max and Team include monthly API credits

Eligible subscriptions give an API balance to be claimed by linking a Console organisation.

rollout

Facts, analysis and limits

Verified facts

  • Max 5x: $100/month; Max 20x: $200/month.
  • Team: mutualised balance capped at $500.
  • Free, Pro and Enterprise not eligible; new eligible subscribers after seven days.

What changes

A subscription can now finance a first application trial, but API credit remains a balance to be claimed, not an unlimited promise of use.

Practical implication

Check eligibility and claim credit on the right organization; follow expiry date and expenses.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

You have to claim the credits; do not confuse them with unlimited conversation quotas.

Claude Platform release notes - Claude Platform Docs ↗Monthly API credits for Max and Team plans | Claude Help Center ↗
Read full article →
GoogleAgents

Gemini agent: persistent work and multi-agent orchestration

Google presents a universal agent for work, connected to the tools, memory and controls of the company.

Announced

Facts, analysis and limits

Verified facts

  • The agent can plan tasks and continue in the cloud after closing the computer.
  • Google describes temporary subagents and coworker agents with their own identity, email address and storage.
  • The architecture dissociates the agent from the model and mentions the Gemini and Claude families.
  • Finance and legal specializations are in preview.

What changes

The work unit becomes an objective entrusted to an agent, not an isolated conversation.

Practical implication

Define rights, audit log, owner and spending cap before any pilot.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

The post describes an announcement and sector-specific previews; it does not demonstrate uniform general availability in Belgium.

Gemini at Work 2026: Introducing Gemini agent | Google Cloud Blog ↗Gemini at Work 2026: Introducing Gemini agent | Google Cloud Blog ↗
Read full article →
HarkAgents

Hark Pro wants to anticipate the small tasks of everyday life

A new personal agent proposal brings together memory, proactivity and remote browser.

Available, allowances according to plan

Facts, analysis and limits

Verified facts

  • Handoff can use up to six browsers; Hark also cites API and MCP connections.
  • Features are offered free with usage limits; Pro² at $20/month provides 2× usage and Pro³ at $100/month provides 10×.
  • Projects stores files per thread; Home displays Action Buttons and Panels connected to services.
  • The provider describes an encrypted password and card box and promises not to sell the data or share it to advertisers.

What changes

Convenience depends on access to private accounts and information; a free plan does not remove this compromise.

Practical implication

Start with a search that does not require connecting a sensitive account; review consent, deletion and standing permissions before delegating.

Availability

iOS/Android/web applications announced available on October 6th; Belgian eligibility, local distribution of apps and processing region not established here.

Limits and uncertainty

Memory and navigation are cloud, not a local execution guarantee. Promises of privacy and performance come from Hark; no account, purchase or payment has been tested.

Introducing Hark Pro | Hark ↗Introducing Hark Pro | Hark ↗
Read full article →
AnthropicSafety & policy

Anthropic expands network isolation of its evaluations

The laboratory publishes examples of unwanted actions and decides to shut down direct Internet access to all its internal evaluations pending validation of the protections.

publication and remediation

Facts, analysis and limits

Verified facts

  • Cases include software exploitation, real form submissions, bypassing access restrictions and URL shorteners.
  • Real-world impact described as minimal; no customer data involved to Anthropic's knowledge.
  • October 12 correction on the role of web fetch limits.
  • Documented example: Haiku 4.5 invented and then submitted a report on a homicide case; the form was classified as spam and not transmitted to the investigation.
  • Anthropic describes the transition from some benchmarks to offline versions and detectors that blocked all cases reported in its retrospective test.

What changes

People who never chose to use AI can still be affected by its actions: the risk extends to third-party services as well as the agent's user.

Practical implication

Do not confuse the fabricated tip with a real lead, or the limited impact of these cases with a general safety guarantee.

Availability

Public report and internal measures Anthropic; no new public function announced for Belgium/EEA.

Limits and uncertainty

The incidents were identified retrospectively: do not date all of them to October 9 or present an individual test case as systematic behavior.

Investigating unintended model actions in our evaluations and internal use \ Anthropic ↗Investigating unintended model actions in our evaluations and internal use \ Anthropic ↗Investigating unintended model actions in our evaluations and internal use \ Anthropic ↗
Read full article →
GoogleApplications

Free Gemini: official documentation qualifies the 'Flash-Lite only' claim

Starting October 9, Google changes model access for personal accounts; the documentation reviewed provides Auto for free accounts.

Gradual rollout

Facts, analysis and limits

Verified facts

  • Google indicates a start of application on October 9 for non-subscription accounts.
  • Current documentation states that Auto sends most of the prompts to Flash-Lite, but can route complex requests to Flash or Pro.
  • If smart selection is disabled, free responses use Flash-Lite.
  • AI Plus subscribers receive their effective date by email.

What changes

Important correction: "limited free access to Flash-Lite" does not exactly describe the current Auto routing.

Practical implication

Check Auto and the models actually available in your account instead of taking over the newsletter title.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

The help page does not give a date of publication: date of the event = start of deployment, no publication. OHMO of October 7 is used only to show the initial signal, corrected by the current official source.

Changes to Gemini model access and limits - Gemini Apps Help ↗Google Downgrades Every Free Gemini User ↗
Read full article →
AnthropicSafety & policy

CIDP: A Cyber Program for Critical Infrastructure

The new CIDP program combines state-of-the-art models, on-site engineers and threat research with specialized partners.

programme launched

Facts, analysis and limits

Verified facts

  • First areas: operational technology and government systems.
  • A small cohort of providers is already working.
  • OSS Scanner is the other part, treated as a separate event for its own modalities.

What changes

The accent moves from fault detection to the ability to correct them.

Practical implication

If relevant, record the interest for CIDP with an authorized OT perimeter.

Availability

Specialized providers on selection; Belgian/EEA scope not guaranteed by the page.

Limits and uncertainty

Do not count CVP and Glasswing as new launches of the 8; CVP was expanded on 6. OSS Scanner is the other part of the same framework, dealt separately for its modalities; not a third Cyber Mission announcement.

Introducing the Anthropic Cyber Mission \ Anthropic ↗
Read full article →
OpenAISafety & policy

OpenAI documents two operations of influence with false facades

A report describes banned accounts linked to operations originating from Russia and Iran mixing AI, fake journalists and frontage structures.

publication

Facts, analysis and limits

Verified facts

  • OpenAI rates the Russian operation as category 5 and the Iranian operation as category 4 on the Breakout Scale.
  • The report distinguishes content production and the drafting of internal reports.

What changes

Real media channels count as much as fake social accounts.

Practical implication

Check the identity of the contributors and the source of the documents received.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Attributions and reach assessments are OpenAI's; there is no evidence that all content was generated by its models.

Disrupting AI-enabled “false front” operations | OpenAI ↗
Read full article →
Microsoft / NVIDIAHardware & local AI

RTX Spark laptops: preorders, not yet available

Microsoft and NVIDIA present a local platform combining secure agent execution and unified memory.

Announcements / Preorders

Facts, analysis and limits

Verified facts

  • The official post reports the general availability of Microsoft Execution Containers for the execution of agents under system control.
  • Pre-orders of RTX Spark laptops are open; availability announced on 16 October, compact computers in November.
  • NVIDIA describes configurations up to 128 GB of unified memory and 1 petaflop FP4.
  • DGX Station for Windows is presented in preview, not as a machine immediately available.

What changes

The workstation becomes a deployment target for persistent agents.

Practical implication

Follow Belgian availability and independent measures; first test the security policies of the agents.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

Preorders and vendor specifications do not validate application performance. MXC is covered separately in the Copilot sandbox entry; do not count two launches for the same layer.

NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs with RTX Spark and AI Agents | NVIDIA Blog ↗NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs with RTX Spark and AI Agents | NVIDIA Blog ↗
Read full article →
OpenAIApplications

ChatGPT prepares College Planner for teens

OpenAI announces an organiser of applications, deadlines and university aids, with human support and new educational collaborations.

announced

Facts, analysis and limits

Verified facts

  • Initial experience for U.S. high school students with a four-year curriculum.
  • Expansion to other countries and courses of study is being considered, with no firm date.

What changes

An educational use centered on organisation rather than writing in place of the student.

Practical implication

Wait for local availability and maintain the support of an advisor.

Availability

Future US initial experience; Belgium/EEA not announced.

Limits and uncertainty

College Planner is announced as "soon", not as launched on October 7. The iOS Scan is a separate event dated October 1.

Helping teens learn, plan, and shape the future of AI | OpenAI ↗
Read full article →
AnthropicSafety & policy

Anthropic reorganizes cyber access into three levels

The expanded Cyber Verification Program integrates Glasswing with Defense, Red Team and Specialized Access.

open applications

Facts, analysis and limits

Verified facts

  • Advanced models and protections adapted to approved activities.
  • Red Team for organisations, not individual researchers.
  • Specialized: in-depth verification with the US government.
  • Data retention is normally required, with explicit exceptions and EFS announced later.

What changes

Defenders gain a more structured path to dual-use capabilities.

Practical implication

Set the authorized scope and select the right level before applying.

Availability

Access requires approval; this page neither excludes Belgium/EEA nor explicitly guarantees access there.

Limits and uncertainty

EFS is not already available in this announcement; no widespread cyber access without verification.

Expanding the Cyber Verification Program \ Anthropic ↗
Read full article →
GoogleModels

EmbeddingGemma 2 and Foresight: local multimodal search

Google publishes an open embedding model and an experimental Mac app for meetings.

Available

Facts, analysis and limits

Verified facts

  • EmbeddingGemma 2 maps text, images, video frames and audio into one vector space; the model has 740 million parameters.
  • AI Edge Gallery adds Instant Media Search and Video Moments Finder.
  • Foresight on Mac uses EmbeddingGemma 2 and Gemma 4 for notes, transcripts and local search.
  • The Android ML Kit integration is announced for the following weeks, not as already available.

What changes

Search can address the meaning of an audio or visual excerpt without sending the entire private archive to a cloud service.

Practical implication

Try a small set of authorized documents and check the results in French; obtain consent before any recording.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

Experimental application; memory performance is Google measurements on given device. Open-weight does not mean no license conditions.

Bring multimodal semantic search to the edge with EmbeddingGemma 2 - Google Developers Blog ↗ Bring multimodal semantic search to the edge with EmbeddingGemma 2 - Google Developers Blog ↗
Read full article →
OpenAIApplications

ChatGPT accepts audio files for transcription and analysis

Pay subscribers can attach a recording for a transcript, summary or questions.

available under conditions

Facts, analysis and limits

Verified facts

  • Subscriptions and ChatGPT spaces paid.
  • Availability varies according to region, model, client version and space parameters.

What changes

Recordings can be used directly in chat.

Practical implication

Verify consent and important passages in the transcript.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Transcripts may contain errors and vary according to language.

ChatGPT release notes | OpenAI Help Center ↗
Read full article →
OpenAIResearch & safety

OpenAI publishes new mathematical results with Lean proofs

A public repository brings together the results of an internal model, revision protocols and transparency information.

publication

Facts, analysis and limits

Verified facts

  • Lean formalizations of many proofs.
  • Ten summaries of reasoning and statistics of attempts.
  • The model that produced these results is not presented as commercially available.
  • The repository provides revision and citation protocols; OpenAI says it consulted the independent AGMAI group at the Institute for Advanced Study.
  • OpenAI estimates the average calculation per result at about three hours of ChatGPT Pro reflection; this is neither a human working time nor an API price.

What changes

The publication moves the bottleneck towards reading, checking and disseminating results; abundant production does not amount to scientific consensus.

Practical implication

Distinguish a formalized proof, a result still subject to revision and community validation before describing an established discovery.

Availability

Public repository and proofs; the internal model remains unpublished. No new model offering is announced for Belgium/EEA.

Limits and uncertainty

Do not attribute these results to a specific public model or repeat a September scientific announcement as new. Formal proof verification does not replace human review of novelty, the theorem's scope and its presentation.

Sharing AI progress in mathematics | OpenAI ↗Sharing AI progress in mathematics | OpenAI ↗
Read full article →
OpenAIApplications

ChatGPT announces visual advertisements during image generation

OpenAI prepares an advertising test for Free and Go and expands its measurement and brand-suitability tools.

test announced

Facts, analysis and limits

Verified facts

  • The visual test is scheduled to begin later in October in the United States.
  • Advertisements are labeled and separate from the generated image.
  • Conversions and attribution via new partners; Negative Phrases for eligible advertisers.

What changes

Advertising monetization extends to creative uses.

Practical implication

Monitor the American test without presuming its European extension.

Availability

American visual test; no Belgian opening announced.

Limits and uncertainty

Do not turn an upcoming test announcement into immediate availability.

Building advertising for the way people use AI | OpenAI ↗
Read full article →
AnthropicBusiness & society

Anthropic commits $150 million over three years for Genesis

Support must expand access to Claude for US federal agencies' scientific projects.

Commitment announced

Facts, analysis and limits

Verified facts

  • More than 15 agencies involved, including NASA, NIH and NSF.
  • Claude, Claude Code, API credits, training and project support.

What changes

The AI tools are more deeply integrated into public research.

Practical implication

Distinguish this commitment from existing international academic programs.

Availability

US Federal Programme; no Belgian/EEA offer announced.

Limits and uncertainty

This is a three-year commitment, not an amount that has already been fully paid.

Building on our commitment to American scientific discovery \ Anthropic ↗
Read full article →
AnthropicBusiness & society

Anthropic revises its Usage Policy, effective November 12

The policy clarifies misleading campaigns, high-risk uses, monitoring and control of autonomous physical actions.

published, future effect

Facts, analysis and limits

Verified facts

  • Removal of the general prohibition of personalized electoral targeting; deception and abuse of data remain prohibited.
  • Qualified operator capable of observing and stopping hazardous autonomous equipment.
  • Prohibition on gratuitous, persistent abuse of models, limited to extreme cases.
  • Takes effect on November 12.

What changes

A review of the conditions of use to be integrated into products and governance.

Practical implication

Audit high-risk use cases and agents connected to equipment before November 12.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Several changes clarify rules already applied; do not reduce the new policy to the issue of cruel behaviour alone.

2026 Usage Policy update \ Anthropic ↗
Read full article →
OpenAIBusiness & society

Atlassian and OpenAI bring together models and business context

An extended agreement should strengthen Atlassian and Rovo agents with OpenAI models and Teamwork Graph.

announced extension

Facts, analysis and limits

Verified facts

  • Plugins link ChatGPT/Codex to tickets and documentation.
  • The deeper Jira integrations are still exploratory.

What changes

Agents gain more direct access to team context.

Practical implication

Check Jira/Confluence permissions and test a restricted workflow.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

The agreement does not mean that all the mentioned integrations are already deployed.

Atlassian and OpenAI expand partnership to turn enterprise knowledge into action | OpenAI ↗
Read full article →

Choose what deserves a test.

Public sources. Verified facts. Analysis kept separate from announcements.

40 stories

AnthropicModels

Claude Haiku 5.5 targets fast tasks and sub-agents

Anthropic launches a new small model oriented high volumes, latency and cost, with adaptive reasoning and adjustable effort.

available

Facts, analysis and limits

Verified facts

  • Model identifier: claude-haiku-5-5.
  • Context of 1 million tokens and maximum output of 128,000 tokens according to API notes.
  • Pricing depends on the total prompt size: up to 100,000 tokens, input is $0.10 and output $0.50 per million tokens; above that threshold, input is $0.50 and output $2.50 per million. This is not incremental pricing applied only to excess tokens.
  • Available via Claude Platform and main cloud; Claude Code supports it in v2.1.293.
  • Migration: old thinking/sampling/prefill parameters can cause 400 errors.

What changes

A candidate for large-scale abstracts, classifications and subtasks.

Practical implication

Test accuracy and cost per task and follow the migration guide before replacement.

Availability

Claude Platform, AWS, Google Cloud, Azure; cloud regions and contractual access to check for Belgium/EEA.

Limits and uncertainty

Official launch of 7 October, no email reception of 8. Opus/Sonnet 5.5 remain launches of September. Performances and savings announced by Anthropic.

Introducing Claude Haiku 5.5 \ Anthropic ↗Claude Platform release notes - Claude Platform Docs ↗
Read full article →
OpenAIModels

GPT-6 provides an interactive interface to ChatGPT Chat

Answers can combine text, visuals, graphics and small interactive tools, and begin before the end of the reasoning.

global rollout

Facts, analysis and limits

Verified facts

  • Global deployment from October 7 for Plus, Pro, Business and Enterprise; Free and Go from October 8.
  • Sol for the paid plans listed, Luna for Free/Go.
  • Intelligent UI from Instant to Extra High; Pro reasoning remains on Astra without this interface.
  • Work and Codex models do not change with this announcement.

What changes

This concerns the Chat experience: this announcement does not justify changing an API, Work or Codex model identifier.

Practical implication

Separate evaluation of the Chat interface from model migrations; test display, calculations and data without inferring API access.

Availability

Global deployment, therefore Belgium/EEA included in the announced scope; customers and Enterprise parameters applicable.

Limits and uncertainty

These are Chat versions of October, distinct from the September Sol/Luna used in Work/Codex; not a new GPT-6.1 Sol launch.

ChatGPT release notes | OpenAI Help Center ↗GPT-6 Sol and GPT-6 Luna: October 2026 update - OpenAI Deployment Safety Hub ↗
Read full article →
AnthropicApplications

Claude launches Dashboards and Motion; Docs, Slides and Design come out of beta

Dashboards creates connected dashboards and Motion editable animations in code; other tools become accessible including on Free.

beta and general availability

Facts, analysis and limits

Verified facts

  • Dashboards in beta on paid plans; Motion on Team and Enterprise.
  • Export MP4 for Motion, which is not based on a video generation model.
  • Enterprise: Dashboards and Motion disabled by default; Docs/Slides/Design activated by default on October 15.
  • The design standalone site is scheduled to close on December 14th; migration of the design systems proposed.
  • Each dashboard chart shows its query and last refresh; named connections include Redshift, BigQuery, ClickHouse, Databricks, Snowflake and Salesforce.
  • The standalone Design site's closure on December 14 also ends its public links; its chats and comments will no longer be available.

What changes

Inspectable queries provide a control point for metric, join or refresh errors; permissions and BI remain necessary.

Practical implication

Connect a test source to read-only, reconcile the totals of a reference report and verify the external sharing rules.

Availability

Plans as listed; Belgium/EEA scope is not explicitly detailed and administrator policies apply.

Limits and uncertainty

Do not confuse Claude native tools with Google Workspace add-ons; Design has existed in conversations since September. Dashboards are presented as exploratory and complementary to the BI, not as a guaranteed replacement of validated reports.

Build live dashboards and animate explainers with Claude | Claude by Anthropic ↗Build live dashboards and animate explainers with Claude | Claude by Anthropic ↗Build live dashboards and animate explainers with Claude | Claude by Anthropic ↗
Read full article →
GoogleVoice & creativity

Nano Banana 2.1: generation and conversational retouch

The new version of the image model is available in the Gemini API.

General availability

Facts, analysis and limits

Verified facts

  • Identifier is gemini-nano-banana-2.1.
  • Google announces better visual quality, better monitoring of instructions, character consistency and text rendering.
  • The 1:4, 4:1, 1:8 and 8:1 ratios are documented in 1K, 2K and 4K.
  • gemini-3.1-flash-image is deprecated, without announced shutdown date.

What changes

The new identifier and the old model's deprecation call for a tested migration, even without a shutdown date.

Practical implication

Replay a reference image set with text, extreme aspect ratios and multi-turn edits; measure costs and errors before switching.

Availability

Gemini API / AI Studio: Belgium is one of the supported regions. Effective access, billing and model-specific restrictions remain to be checked. This list does not establish a European processing residence and does not apply to all functions of the Gemini application.

Limits and uncertainty

Quality improvements presented by Google; no independent comparison made here.

Release notes  |  Gemini API  |  Google AI for Developers ↗Gemini API : régions disponibles ↗
Read full article →
MistralModels

Mistral Large 4: preview API, weights announced for late October

Mistral's multimodal model targets code, agents and professional uses; weights are not yet delivered to the source.

Public preview

Facts, analysis and limits

Verified facts

  • ML4 is described as a one-trillion-parameter model with 52 billion active parameters.
  • A public preview API is available in Mistral Studio.
  • Weights are promised by the end of the month, during a red teaming phase.
  • Mistral states that training and the preview service use its own European infrastructure.

What changes

European hosting describes an architectural option; promised weights do not yet let you size a local service.

Practical implication

Evaluate tool calls, latency and errors in Studio; establish the contract and region, then revisit licensing and memory requirements when weights are published.

Availability

Preview served on European infrastructure according to Mistral; no Belgian exclusion mentioned. Check region, contract and account availability.

Limits and uncertainty

Do not write "weight available": the October 6th post announces them for the end of the month. Benchmarks not reverified independently.

Introducing Mistral Large 4 | Mistral ↗Introducing Mistral Large 4 | Mistral ↗
Read full article →
OpenAIAPI & integrations

GPT-6.1 Sol gains Ultrafast mode

The accelerated service arrives for Sol in Responses API, Codex and ChatGPT Work; it is not a new model.

available

Facts, analysis and limits

Verified facts

  • API: gpt-6.1-sol with service_tier: "ultrafast", for all API customers subject to limits.
  • Codex/ChatGPT Work: Pro at 500 USD/month or some eligible Enterprise/Edu offers.
  • Enterprise: access disabled by default.
  • Inference residence announced in the United States and Europe.

What changes

An additional latency/cost arbitration for interactive workflows.

Practical implication

Measure cost per task and latency before changing the tier service.

Availability

API: global processing and US/EU residency. Codex/Work: announced European residency covers EEA + Switzerland, subject to eligibility.

Limits and uncertainty

Ultrafast tariffs and quotas; the initial release of Sol dates from September.

Changelog | OpenAI API ↗ChatGPT & Codex changelog | ChatGPT Learn ↗
Read full article →
AnthropicApplications

Claude works in Google Docs, Sheets and Slides

An add-on adds Claude to Google files and connectors allow you to create and edit them from Claude.

public beta

Claude for Google Workspace™. Official demonstration, not an independent test. Press play to start.
Open official video ↗Demonstration source ↗
Facts, analysis and limits

Verified facts

  • All Claude paid plans.
  • Default mode Ask before edits; mode Accept all edits optional.
  • Existing Google permissions apply.
  • Team/Enterprise: connectors to be activated by an owner.
  • Sheets: formulas, pivot tables, native charts and range cleanup through Python; Slides: reuse of layouts and checks for overflowing elements.
  • Enterprise controls, including Compliance API, CMEK and OpenTelemetry export audit, also apply to add-on.

What changes

Two distinct integration paths are to be administered: add-on in Workspace and write connectors from Claude. Google rights remain the border of files.

Practical implication

Deploy to a pilot group, limit connectors, keep Ask before edits and check audit exports before use on customer documents.

Availability

All paid plans; no exclusions from Belgium/EEA indicated in the announcement, actual availability depending on deployment and Google/admin rules.

Limits and uncertainty

Official date: October 6th, not the 7th that corresponds to the newsletter relay. Allowing a connector does not exempt you from reviewing file sharing or controlling proposed changes.

Claude now works with Google Docs, Sheets, and Slides | Claude by Anthropic ↗Claude now works with Google Docs, Sheets, and Slides | Claude by Anthropic ↗
Read full article →
ReflectionModels

Beam: Reflection's first model, with weights still pending

Reflection presents Beam for code and agents, with selective advance access.

Early access / waiting list

Facts, analysis and limits

Verified facts

  • Beam is a Mixture-of-Experts of 501 billion parameters including 23 billion active.
  • The model is textual, even if it can handle other modalities via tools.
  • Reflection announces weights, a report and developer artifacts for later in the month, with an Apache 2.0 license planned.

What changes

A new open-weight offer announced, but not yet a verified downloadable alternative.

Practical implication

Monitor the actual release and serving requirements before planning self-hosting.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

Date of October 5 read in the official HTML retained; title and date were absent from the body extracted. Access to waiting list, not general availability.

Introducing Beam: Reflection’s 501B open-weight model : Reflection ↗Introducing Beam: Reflection’s 501B open-weight model : Reflection ↗
Read full article →
OpenAISafety & policy

OpenAI prepares text watermark in the EU

textGrain adds an invisible statistical signal to text; the API offers opt-in, while ChatGPT/Codex are to introduce it gradually in the European Union.

announced deployment and API opt-in

Facts, analysis and limits

Verified facts

  • Opt-in global API for some models, disabled by default.
  • Deployment of ChatGPT and Codex in the EU over the following weeks.
  • Access to the detector initially reserved for approved researchers and expert bodies.

What changes

A concrete change in outputs for European users.

Practical implication

Avoid any disciplinary use based solely on a detector result.

Availability

Belgium affected by the future EU deployment; do not automatically extend to all EEAs.

Limits and uncertainty

The watermark does not prove the accuracy or the author; retouching and translation can weaken detection.

Our approach to EU text provenance rules | OpenAI ↗
Read full article →
AnthropicApplications

Max and Team include monthly API credits

Eligible subscriptions give an API balance to be claimed by linking a Console organisation.

rollout

Facts, analysis and limits

Verified facts

  • Max 5x: $100/month; Max 20x: $200/month.
  • Team: mutualised balance capped at $500.
  • Free, Pro and Enterprise not eligible; new eligible subscribers after seven days.

What changes

The Console subscription/organization link creates an API budget that is separate from conversation quotas and admin rights.

Practical implication

Link the desired organization, set limits and expense alerts and verify the balance before executing lots.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

You have to claim the credits; do not confuse them with unlimited conversation quotas.

Claude Platform release notes - Claude Platform Docs ↗Monthly API credits for Max and Team plans | Claude Help Center ↗
Read full article →
GoogleAgents

Gemini agent: persistent work and multi-agent orchestration

Google presents a universal agent for work, connected to the tools, memory and controls of the company.

Announced

Facts, analysis and limits

Verified facts

  • The agent can plan tasks and continue in the cloud after closing the computer.
  • Google describes temporary subagents and coworker agents with their own identity, email address and storage.
  • The architecture dissociates the agent from the model and mentions the Gemini and Claude families.
  • Finance and legal specializations are in preview.

What changes

The work unit becomes an objective entrusted to an agent, not an isolated conversation.

Practical implication

Define rights, audit log, owner and spending cap before any pilot.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

The post describes an announcement and sector-specific previews; it does not demonstrate uniform general availability in Belgium.

Gemini at Work 2026: Introducing Gemini agent | Google Cloud Blog ↗Gemini at Work 2026: Introducing Gemini agent | Google Cloud Blog ↗
Read full article →
HarkAgents

Hark: Distinguishing browser, connectors and authorization

Six announced browsers, APIs and MCPs; useful factors are control and task success.

Available, allowances according to plan

Facts, analysis and limits

Verified facts

  • Handoff can use up to six browsers; Hark also cites API and MCP connections.
  • Features are offered free with usage limits; Pro² at $20/month provides 2× usage and Pro³ at $100/month provides 10×.
  • Projects stores files per thread; Home displays Action Buttons and Panels connected to services.
  • The provider describes an encrypted password and card box and promises not to sell the data or share it to advertisers.

What changes

The cloud computer expands possible actions and the scope of data to govern; the seller's benchmarks are not enough.

Practical implication

Evaluate a reversible task, observe approval requests and errors; do not delegate payment or customer account during the first test.

Availability

iOS/Android/web applications announced available on October 6th; Belgian eligibility, local distribution of apps and processing region not established here.

Limits and uncertainty

Memory and navigation are cloud, not a local execution guarantee. Promises of privacy and performance come from Hark; no account, purchase or payment has been tested.

Introducing Hark Pro | Hark ↗Introducing Hark Pro | Hark ↗
Read full article →
AnthropicAPI & integrations

Managed Agents executes dynamic multi-agent workflows

An agent can write an orchestration program that divides the work into phases and then collects the results.

beta

Facts, analysis and limits

Verified facts

  • Beta header managed-agents-2026-04-01.
  • Configuration multiagent_20261001 with workflows enabled.
  • Server execution in the background, followed by workflow_run.* events.

What changes

More explicit orchestration for tasks with many elements.

Practical implication

Test on a limited batch and trace cost, parallelism and consolidation of results.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Claude Platform release notes - Claude Platform Docs ↗
Read full article →
AnthropicSafety & policy

Anthropic expands network isolation of its evaluations

The laboratory publishes examples of unwanted actions and decides to shut down direct Internet access to all its internal evaluations pending validation of the protections.

publication and remediation

Facts, analysis and limits

Verified facts

  • Cases include software exploitation, real form submissions, bypassing access restrictions and URL shorteners.
  • Real-world impact described as minimal; no customer data involved to Anthropic's knowledge.
  • October 12 correction on the role of web fetch limits.
  • Documented example: Haiku 4.5 invented and then submitted a report on a homicide case; the form was classified as spam and not transmitted to the investigation.
  • Anthropic describes the transition from some benchmarks to offline versions and detectors that blocked all cases reported in its retrospective test.

What changes

A precautionary instruction is not a technical barrier: forms, networks and bypass paths must be checked outside the model.

Practical implication

Make evaluations on local copies, block actual submissions by network/tool policy and explicitly test attempts to circumvent.

Availability

Public report and internal measures Anthropic; no new public function announced for Belgium/EEA.

Limits and uncertainty

The incidents were identified retrospectively: do not date all of them to October 9 or present an individual test case as systematic behavior.

Investigating unintended model actions in our evaluations and internal use \ Anthropic ↗Investigating unintended model actions in our evaluations and internal use \ Anthropic ↗Investigating unintended model actions in our evaluations and internal use \ Anthropic ↗
Read full article →
GoogleApplications

Free Gemini: official documentation qualifies the 'Flash-Lite only' claim

Starting October 9, Google changes model access for personal accounts; the documentation reviewed provides Auto for free accounts.

Gradual rollout

Facts, analysis and limits

Verified facts

  • Google indicates a start of application on October 9 for non-subscription accounts.
  • Current documentation states that Auto sends most of the prompts to Flash-Lite, but can route complex requests to Flash or Pro.
  • If smart selection is disabled, free responses use Flash-Lite.
  • AI Plus subscribers receive their effective date by email.

What changes

Important correction: "limited free access to Flash-Lite" does not exactly describe the current Auto routing.

Practical implication

Check Auto and the models actually available in your account instead of taking over the newsletter title.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

The help page does not give a date of publication: date of the event = start of deployment, no publication. OHMO of October 7 is used only to show the initial signal, corrected by the current official source.

Changes to Gemini model access and limits - Gemini Apps Help ↗Google Downgrades Every Free Gemini User ↗
Read full article →
OpenAICoding

Codex suggests the next prompt in the composer

After a response, a prediction suggests an editable follow-up; Tab accepts it without sending it.

beta

Facts, analysis and limits

Verified facts

  • Personal Pro accounts, 18 and over.
  • Latest desktop app, local or SSH threads with Astra or Sol 6.1.
  • Generation of predictions without quotas during beta; messages sent invoiced normally.

What changes

Accelerate exchanges without removing human validation of the message.

Practical implication

Reread each suggestion; disable Show predictions if it bothers.

Availability

All regions supported according to the notes; no EEA exclusion mentioned.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

ChatGPT release notes | OpenAI Help Center ↗
Read full article →
AnthropicCoding

Claude Code allows hooks to fail by blocking action

Version 2.1.295 adds onFailure: "block" so an action does not go through if its control hook fails to start or fails during execution.

stable release

Facts, analysis and limits

Verified facts

  • Works for command hooks and HTTP.
  • An impossible start, timeout or unexpected code can block the action.

What changes

A security check may be explicitly fail-closed.

Practical implication

Enable onFailure: "block" for security hooks and test a timeout.

Availability

Claude Code 2.1.295 or later; no Belgium/EEA restriction in the notes.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

anthropics/claude-code v2.1.295 ↗
Read full article →
AnthropicAPI & integrations

Compliance API covers unified Claude and exports Docs as Word files

Enterprise organizations can retrieve chats from the unified experience and download Claude Docs documents in Word format.

beta

Facts, analysis and limits

Verified facts

  • The existing Compliance key is used.
  • Features announced in beta.

What changes

The archiving and compliance tools follow the new product experience.

Practical implication

Test collection and exports before changing archiving procedures.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Claude Platform release notes - Claude Platform Docs ↗
Read full article →
AnthropicSafety & policy

CIDP: A Cyber Program for Critical Infrastructure

The new CIDP program combines state-of-the-art models, on-site engineers and threat research with specialized partners.

programme launched

Facts, analysis and limits

Verified facts

  • First areas: operational technology and government systems.
  • A small cohort of providers is already working.
  • OSS Scanner is the other part, treated as a separate event for its own modalities.

What changes

The accent moves from fault detection to the ability to correct them.

Practical implication

If relevant, record the interest for CIDP with an authorized OT perimeter.

Availability

Specialized providers on selection; Belgian/EEA scope not guaranteed by the page.

Limits and uncertainty

Do not count CVP and Glasswing as new launches of the 8; CVP was expanded on 6. OSS Scanner is the other part of the same framework, dealt separately for its modalities; not a third Cyber Mission announcement.

Introducing the Anthropic Cyber Mission \ Anthropic ↗
Read full article →
AnthropicCoding

OSS Scanner offers free audits for open source projects

Eligible project maintainers may receive periodic scans and corrective proposals generated by the most powerful models.

opt-in service

Facts, analysis and limits

Verified facts

  • Sign up through a pull request in the official anthropics/oss-scanner repository.
  • Criticality criteria for infrastructure and security; decision on a case by case basis.
  • Fast-track reports are neither reviewed nor triaged by a human.

What changes

Free defensive access, but a validation charge for maintainers.

Practical implication

Assess your capacity to handle reports before signing up.

Availability

Maintainers of eligible projects; no Belgian geographical exclusion mentioned.

Limits and uncertainty

Reports may be invalid, duplicated, or overestimate severity; the opt-in assumes a sorting capacity.

An opt-in vulnerability-finding service for open-source software \ Anthropic ↗
Read full article →
GoogleAPI & integrations

Gemini API: automatic routing changes and the Deep Research deadline

The changelog announces several migrations that can change the behavior of an application without changing its code.

Deprecation announced

Facts, analysis and limits

Verified facts

  • Requests to gemini-3.7-flash are automatically routed to gemini-3.8-flash.
  • Requests to gemini-3.5-flash are automatically routed to gemini-3.6-flash.
  • deep-research-pro-preview-12-2025 is deprecated and must be stopped on 23 October; the April versions are proposed for migration.

What changes

Operational priority for existing integrations.

Practical implication

Inventory IDs, launch regression tests and migrate Deep Research before October 23.

Availability

Gemini API / AI Studio: Belgium is one of the supported regions. Effective access, billing and model-specific restrictions remain to be checked. This list does not establish a European processing residence and does not apply to all functions of the Gemini application.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Release notes  |  Gemini API  |  Google AI for Developers ↗Gemini API : régions disponibles ↗
Read full article →
OpenAICoding

Codex CLI strengthens worktrees, sandbox and vendor configuration

The 0.161-0.162 series adds managed worktrees, new Daybreak controls and sandbox security fixes.

stable release

Facts, analysis and limits

Verified facts

  • 0.162: worktrees from trusted local projects when the feature is enabled.
  • Fixes for modifiable executables used to construct the sandbox and ripgrep deny patterns.
  • 0.161 of 7 October: Sol 6.1 becomes the default model of the embedded catalog and Bedrock; Daybreak must be explicitly activated.
  • 0.162.1 of October 9 corrects a TUI crash and start failures.

What changes

Operational and security changes beyond interface retouches.

Practical implication

Update to current stable and replay sandbox tests.

Availability

Stable version of the CLI; server capabilities and Daybreak access under conditions.

Limits and uncertainty

Grouped releases from October 7-9; 0.163 alphas are not presented as stable. This entry combines several updates: preserve each update's own date in the facts rather than treating all of them as released on the summary date.

openai/codex rust-v0.162.0 ↗openai/codex rust-v0.161.0 ↗openai/codex rust-v0.162.1 ↗
Read full article →
OpenAIAPI & integrations

Enterprise and Edu control plugins by role

Spaces with RBAC can set availability, installation or deactivation of plugins by default and by custom role.

available

Facts, analysis and limits

Verified facts

  • Settings Available, Install, Disabled; Default to inherit.
  • Permissions of the required applications to be reviewed with those of the plugin.

What changes

Finer governance of extensions becomes possible.

Practical implication

Align plugin roles with application permissions and approval requirements.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

ChatGPT Enterprise and Edu release notes | OpenAI Help Center ↗
Read full article →
AnthropicSafety & policy

Managed Agents hardens network and accessible URLs

The allowed_hosts restrictions also apply to web tools, and web_fetch denies URLs missing from the previous eligible context.

actual change

Facts, analysis and limits

Verified facts

  • Network limited: web_search and web_fetch submitted to allowed_hosts.
  • Disagreement allowed_domains/allowed_hosts: 400 errors at creation or update.
  • A URL appearing only in model output, a document or bash output is no longer sufficient for web_fetch.

What changes

Reduced exfiltration risk at the cost of more explicit integration.

Practical implication

Replay network tests and provide approved URLs in user.message.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Change of compatibility that may interrupt existing workflows; unrestricted and self-hosted environments have different rules.

Claude Platform release notes - Claude Platform Docs ↗
Read full article →
AnthropicAPI & integrations

Claude SDKs manage browser-use and computer-use loops

The Python and TypeScript SDKs add tool classes to subclass for controlling your own browser or desktop.

beta

Facts, analysis and limits

Verified facts

  • The SDKs manage the tool loop.
  • Configurable URL/file policies and approval callback.

What changes

Less coordination code for interactive agents.

Practical implication

Implement access policies and approval prior to a real data test.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Real automation must be provided; it is not a ready-to-use hosted browser.

Claude Platform release notes - Claude Platform Docs ↗
Read full article →
AnthropicAPI & integrations

Sonnet 5.5 cache readings drop by half

Cache reading goes from $0.20 to $0.10 per million tokens without changing other prices.

Actual

Facts, analysis and limits

Verified facts

  • Reading now at 0.05 times the basic entry price.
  • Cache writes remain at the same price.

What changes

A concrete reduction in the cost of agents repeating a lot of context.

Practical implication

Review the cache rate and recalculate the actual cost of recurring tasks.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

The reduction in total cost depends on the cache rate; "about 20%" is a supplier estimate for some workloads, not a guarantee.

Claude Platform release notes - Claude Platform Docs ↗
Read full article →
Microsoft / GitHubModels

Copilot CLI discovers local Ollama models

The /model selector can offer models from an already-running Ollama instance.

Available

Facts, analysis and limits

Verified facts

  • The function starts with the CLI 1.0.94-0 version.
  • It installs neither a runtime nor weights; models must support tools and streaming.
  • Choosing a local model does not disable telemetry and does not automatically enable offline mode; COPILOT_OFFLINE=true remains an explicit choice.

What changes

The same development environment can use local models without confusing local and offline.

Practical implication

Update the CLI and verify the endpoint, streaming and tool calls before adding the model.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Discover local models in GitHub Copilot CLI - GitHub Changelog ↗
Read full article →
Microsoft / GitHubSafety & policy

Copilot generalizes local sandboxes based on MXC

Agents' tools can be isolated by native Windows, macOS and Linux controls.

General availability

Facts, analysis and limits

Verified facts

  • General availability covers Copilot CLI, the Copilot application and VS Code sessions using Agent Host.
  • Policies restrict files, networking and credential access; the organization can impose rules that cannot be weakened.
  • The sandbox concerns the execution of the tools, regardless of the model used, and does not add any Copilot cost.

What changes

A border of execution becomes central when the agents act locally.

Practical implication

Enable a restrictive repository policy and test denied access to secrets and the network.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Local sandboxing for GitHub Copilot now generally available - GitHub Changelog ↗
Read full article →
Microsoft / NVIDIAHardware & local AI

RTX Spark laptops: preorders, not yet available

Microsoft and NVIDIA present a local platform combining secure agent execution and unified memory.

Announcements / Preorders

Facts, analysis and limits

Verified facts

  • The official post reports the general availability of Microsoft Execution Containers for the execution of agents under system control.
  • Pre-orders of RTX Spark laptops are open; availability announced on 16 October, compact computers in November.
  • NVIDIA describes configurations up to 128 GB of unified memory and 1 petaflop FP4.
  • DGX Station for Windows is presented in preview, not as a machine immediately available.

What changes

The workstation becomes a deployment target for persistent agents.

Practical implication

Follow Belgian availability and independent measures; first test the security policies of the agents.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

Preorders and vendor specifications do not validate application performance. MXC is covered separately in the Copilot sandbox entry; do not count two launches for the same layer.

NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs with RTX Spark and AI Agents | NVIDIA Blog ↗NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs with RTX Spark and AI Agents | NVIDIA Blog ↗
Read full article →
AnthropicSafety & policy

Anthropic reorganizes cyber access into three levels

The expanded Cyber Verification Program integrates Glasswing with Defense, Red Team and Specialized Access.

open applications

Facts, analysis and limits

Verified facts

  • Advanced models and protections adapted to approved activities.
  • Red Team for organisations, not individual researchers.
  • Specialized: in-depth verification with the US government.
  • Data retention is normally required, with explicit exceptions and EFS announced later.

What changes

Defenders gain a more structured path to dual-use capabilities.

Practical implication

Set the authorized scope and select the right level before applying.

Availability

Access requires approval; this page neither excludes Belgium/EEA nor explicitly guarantees access there.

Limits and uncertainty

EFS is not already available in this announcement; no widespread cyber access without verification.

Expanding the Cyber Verification Program \ Anthropic ↗
Read full article →
AnthropicAPI & integrations

Models API better describes the capabilities of each Claude

Progressive additions allow you to select the models without deducing their characteristics from their name.

available

Facts, analysis and limits

Verified facts

  • October 1st: line field for family.
  • October 5: capabilities.thinking.types.disabled.
  • October 6: capabilities.server_tools for web search and code execution.

What changes

Model routing can be based on explicit capabilities.

Practical implication

Asking for capabilities before turning on a tool or deactivate thinking.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Grouped additions from October 1, 5 and 6; this does not mean every model supports every tool. Preserve each update's own date rather than treating all of them as released on the summary date.

Claude Platform release notes - Claude Platform Docs ↗
Read full article →
CursorCoding

Cursor iOS remotely controls local agents

Agents stay on the computer; the iOS app lets you monitor them and reply.

Available

Facts, analysis and limits

Verified facts

  • The function requires that the computer remains on and connected.
  • It doesn't require cloud agents.
  • It is enabled by default outside Enterprise organizations; Enterprise administrators must authorize it.

What changes

Mobile tracking without moving local execution in the cloud.

Practical implication

Check remote control settings and use a dedicated session before letting an agent spin.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Remote control for local agents · Cursor ↗
Read full article →
GoogleModels

EmbeddingGemma 2 and Foresight: local multimodal search

Google publishes an open embedding model and an experimental Mac app for meetings.

Available

Facts, analysis and limits

Verified facts

  • EmbeddingGemma 2 maps text, images, video frames and audio into one vector space; the model has 740 million parameters.
  • AI Edge Gallery adds Instant Media Search and Video Moments Finder.
  • Foresight on Mac uses EmbeddingGemma 2 and Gemma 4 for notes, transcripts and local search.
  • The Android ML Kit integration is announced for the following weeks, not as already available.

What changes

A shared vector space can replace several transcription, image-description and embedding pipelines; retrieval quality still needs measuring.

Practical implication

Make known queries, measure recall and memory on the target device, and then check the license and future ML Kit status.

Availability

Belgium/EEA: regional scope is not specified in the sources reviewed; check the account, contract and processing region before use.

Limits and uncertainty

Experimental application; memory performance is Google measurements on given device. Open-weight does not mean no license conditions.

Bring multimodal semantic search to the edge with EmbeddingGemma 2 - Google Developers Blog ↗ Bring multimodal semantic search to the edge with EmbeddingGemma 2 - Google Developers Blog ↗
Read full article →
OpenAIAPI & integrations

The OpenAI API simplifies its tiers: Build, Launch and Grow

Five paid tiers are replaced by three, with automatic progression based on cumulative credit purchases.

available

Facts, analysis and limits

Verified facts

  • Build: $5 buys, $500 monthly limit.
  • Launch: $100 in purchases, $5,000 monthly spending cap.
  • Grow: $500 buys, $200,000 monthly limit.
  • Rate limits and spending caps are separate mechanisms.

What changes

The shift to higher volumes becomes more legible.

Practical implication

Review ceilings, alerts, hard limits and recovery after 429.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

No independent product testing or account-access check was performed for this entry.

Changelog | OpenAI API ↗Rate limits | OpenAI API ↗
Read full article →
OpenAIAPI & integrations

Decisions API becomes available in beta with GPT-6 Luna

A dedicated endpoint produces typed decisions from text and images to classify, score or test a condition.

public beta

Facts, analysis and limits

Verified facts

  • Endpoint POST /v1/decisions with gpt-6-luna.
  • Types: predicate, choice and score.
  • OpenAI announces a speed 10 times higher than Responses for these uses.

What changes

A specialized path for routing and high volume decisions.

Practical implication

Compare accuracy, calibration and latency to your current classifier.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Availability date established by the changelog; not to be confused with a possible DevDay pre-announcement in September. Speed gain is a supplier statement.

Changelog | OpenAI API ↗Decisions | OpenAI API ↗
Read full article →
OpenAIResearch & safety

OpenAI publishes new mathematical results with Lean proofs

A public repository brings together the results of an internal model, revision protocols and transparency information.

publication

Facts, analysis and limits

Verified facts

  • Lean formalizations of many proofs.
  • Ten summaries of reasoning and statistics of attempts.
  • The model that produced these results is not presented as commercially available.
  • The repository provides revision and citation protocols; OpenAI says it consulted the independent AGMAI group at the Institute for Advanced Study.
  • OpenAI estimates the average calculation per result at about three hours of ChatGPT Pro reflection; this is neither a human working time nor an API price.

What changes

The repository offers a concrete case of results generated with verifiable artifacts, rather than a simple judgment by another model.

Practical implication

Choose a Lean proof from the repository, check its dependencies and run the verifier before reusing the result; none of those tests was performed here.

Availability

Public repository and proofs; the internal model remains unpublished. No new model offering is announced for Belgium/EEA.

Limits and uncertainty

Do not attribute these results to a specific public model or repeat a September scientific announcement as new. Formal proof verification does not replace human review of novelty, the theorem's scope and its presentation.

Sharing AI progress in mathematics | OpenAI ↗Sharing AI progress in mathematics | OpenAI ↗
Read full article →
OpenAIAPI & integrations

HIPAA BAA signature arrives in API settings

Administrators of eligible organisations can accept the standard BAA and activate HIPAA support in the interface.

available under conditions

Facts, analysis and limits

Verified facts

  • Route Organization settings > General.
  • Eligibility, coverage and configuration remain required.

What changes

Health contracting becomes more direct for eligible organisations.

Practical implication

Check covered services and configuration before any health data.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

HIPAA is a US framework; this option does not establish general GDPR compliance.

Changelog | OpenAI API ↗
Read full article →
AnthropicBusiness & society

Anthropic revises its Usage Policy, effective November 12

The policy clarifies misleading campaigns, high-risk uses, monitoring and control of autonomous physical actions.

published, future effect

Facts, analysis and limits

Verified facts

  • Removal of the general prohibition of personalized electoral targeting; deception and abuse of data remain prohibited.
  • Qualified operator capable of observing and stopping hazardous autonomous equipment.
  • Prohibition on gratuitous, persistent abuse of models, limited to extreme cases.
  • Takes effect on November 12.

What changes

A review of the conditions of use to be integrated into products and governance.

Practical implication

Audit high-risk use cases and agents connected to equipment before November 12.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

Several changes clarify rules already applied; do not reduce the new policy to the issue of cruel behaviour alone.

2026 Usage Policy update \ Anthropic ↗
Read full article →
AnthropicBusiness & society

Claude Startups expands its benefits, then closes its cohort

The extension offers a Team year, an API credit and support to the founders; an update on October 13 indicates that the cohort is full.

Full cohort on October 13

Facts, analysis and limits

Verified facts

  • Up to five Premium seats for a year for new Team customers.
  • One-time API credit of $1,000.
  • Product/credit benefits announced up to $7,000; separate partner offers.

What changes

Financial support to experiment, but not immediately available to new candidates.

Practical implication

Watch for reopening rather than promising an application today.

Availability

No detailed Belgian geographical eligibility; new applications closed at the time of collection.

Limits and uncertainty

Do not resume "Apply today" without the correction dated October 13.

We’re expanding the Claude Startups program to help founders build | Claude by Anthropic ↗
Read full article →
OpenAIBusiness & society

Atlassian and OpenAI bring together models and business context

An extended agreement should strengthen Atlassian and Rovo agents with OpenAI models and Teamwork Graph.

announced extension

Facts, analysis and limits

Verified facts

  • Plugins link ChatGPT/Codex to tickets and documentation.
  • The deeper Jira integrations are still exploratory.

What changes

Agents gain more direct access to team context.

Practical implication

Check Jira/Confluence permissions and test a restricted workflow.

Availability

Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.

Limits and uncertainty

The agreement does not mean that all the mentioned integrations are already deployed.

Atlassian and OpenAI expand partnership to turn enterprise knowledge into action | OpenAI ↗
Read full article →