Journal IA 2026-W40

Verified facts

• Google announces Gemini 4 Argon on September 30 and starts a progressive access for trusted cyber defenders via Fairwind.
• The announcement targets long-term professional tasks: development, finance, law and cybersecurity.
• The general availability of 13 October is the subject of a separate sheet; do not confuse this initial announcement with the subsequent deployment.

What changes

A new stage of competition on complex workflows, to be distinguished from product availability.

Practical implication

Read also the October 13 update before deducing what is accessible.

Availability

As of September 30: limited progressive access, no general availability. See the October 13 update for API and application.

Limits and uncertainty

Retrospective addition: this announcement was missed in the pilot. Performance claims in the post are Google's and were not reproduced here.

Verified facts

• Dots is an always-on agent powered by GPT-6 Astra, with its own cloud computer, that can pursue a goal between conversations.
• The Pro rollout excludes the EEA, Switzerland and UK at launch; Enterprise access is a beta disabled by default.
• Usage temporarily does not count against the allocation during the first month for eligible plans; subsequent terms are not yet detailed in the notes reviewed.

What changes

The change to be monitored is the ongoing delegation: preparation, research and follow-up between two conversations.

Practical implication

Watch for European access; do not switch plans for Dots from Belgium.

Availability

Belgium: no Pro access to launch. Do not extrapolate this restriction to professional offers without checking their conditions.

Limits and uncertainty

An agent available continuously is not a guarantee of correct result. Authorizations and review remain necessary.

Verified facts

• GPT-6.1 Sol succeeds GPT-6 Sol for agent coding, computer use and professional work; proximity to Astra is an assertion of OpenAI, not a test performed for this journal.
• For standard API requests with up to 272,000 input tokens: $2 per million input tokens, $0.10 for cache reads and $10 for output; above the threshold, long-context rates apply to the entire request.
• The announced window is 1,050,000 tokens. Tool calls use Responses; Chat Completions is supported without tool calls.
• Supported effort levels are low, medium, high, xhigh and max. EU data residency is supported, but Fast mode is not supported with that residency.

What changes

Useful progression is not only a brighter answer: it is the ability to carry out a task with tools until the result.

Practical implication

Test on a known task before replacing your default model.

Availability

API available; Work/Codex gradually deployed starting with Pro. API subscription and billing are two distinct channels.

Limits and uncertainty

The maximum context is not free, and the price per token does not measure the cost of a complete task.

Verified facts

• Opus 5.5 is announced on September 22nd and Sonnet 5.5 on 28. Both are available on the Claude platform and the main clouds mentioned by Anthropic.
• API prices per million tokens: Opus 5.5 at $4 input / $20 output; Sonnet 5.5 at $2 / $10. Claude subscriptions have their own limits.
• Anthropic announces more than 30% additional generation speed for both; the announced savings per task are 40% for Opus and up to 30% for Sonnet, on its own tests.
• Five-hour limits increase on plans including Pro and Max with Opus 5.5; a limit reset can be saved and used later. Haiku 5.5 is announced for later, not available in these announcements.

What changes

More efficient models can make the same use allocation more useful, without requiring an increase in budget.

Practical implication

Check the new Max limits and compare on your usual usages.

Availability

Available; increased limits announced for several subscriptions including Max. No status of your personal account has been inspected.

Limits and uncertainty

Efficiency figures are the vendor's. Le Signal AI is a practitioner's account, not proof that Opus wins on every task.

Sonnet update September 28; reminder of Opus launch on 22. Both dates are distinct.

Verified facts

• MAI-Transcribe-2-Streaming handles 60 languages with continuous language detection and initial hypotheses in just over 100 ms, according to Microsoft.
• The advertised introductory price is $0.54 per audio hour until the end of the year.
• MAI-Voice-2.1 and Flash support 23 languages and 26 regional variants, with announced rates of $22 and $15 per million characters.
• The three models are accessible via Microsoft Foundry and MAI Playground; LiveKit is indicated as coming.

What changes

The partial transcript may allow the agent to prepare his answer before the end of the sentence.

Practical implication

Measure latency end-to-end, errors in French and consent before clone a voice.

Availability

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

Limits and uncertainty

Rankings and latencies cited come from Microsoft's post; check regional availability and EEA residency.

Verified facts

• The official website offers ESP32 and Linux SDKs and examples for Raspberry Pi, screens and sensors.
• The firmware and SDKs are under Apache 2.0 license and provided without warranty.
• The Muse Home Link offer is reserved for Muse subscribers in the United States; the application is not an order.

What changes

An open hardware layer can extend agents beyond the browser.

Practical implication

Use a dedicated development card without opening the sensitive home network.

Availability

Open code for experimentation; access to the Muse service from Belgium is not established. Home Link is explicitly US-only.

Limits and uncertainty

October 2 is attested by TechCrunch; the official site reviewed is undated. Opening the SDKs does not mean opening the Muse model.

Verified facts

• The configuration retains GB10, DGX OS and the NVIDIA AI stack.
• Two 64 GB units can be connected and managed via Sync Cluster Assistant.
• The partner availability announced is October 23, starting at $4,999.
• The Model Launcher Sync, including an OpenCode configuration, is announced by the end of the month.

What changes

A new local service option; the date of announcement should not be confused with the shipment.

Practical implication

Size memory and throughput for your actual model before considering a purchase.

Availability

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

Limits and uncertainty

Price announced in dollars; Belgian price, VAT and availability unverified reseller. Performance gains specific to NVIDIA tests.

Verified facts

• Web, iOS and Android.
• Explicitly American deployment.

What changes

Specialized functionality reaches a wider audience.

Practical implication

Do not present it as available in Belgium.

Availability

United States only in this announcement; Belgium/EEA not announced.

Limits and uncertainty

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

Verified facts

• Camera > three-point menu > Scan.
• Deployment to iOS.

What changes

Less friction to use paper notes and documents.

Practical implication

Test with non-confidential notes and check the OCR.

Availability

iOS; Belgian availability not detailed. Android is not announced as available.

Limits and uncertainty

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

Verified facts

• Try on button on clothing and accessory product cards.
• The reference photo is preserved, editable and removeable in Customization.
• Favorites and product folders on mobile and web.

What changes

Shopping becomes a personalized visual experience.

Practical implication

Check the settings of the reference photos before testing.

Availability

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

Limits and uncertainty

Countries and plans not detailed in this entry.

Verified facts

• Anthropic's September 29 study describes advanced cyber-exploitation capabilities in GLM-5.3 and safeguards that can be bypassed in its simulations.
• NIST/CAISI ranks GLM-5.3 as the most cyber-capable open-weight model among those released at that date, but behind the US frontier on its aggregate measure.
• The NIST dates the release of GLM-5.3 on August 14th and the release of the weights two weeks later: the news of this fortnight is the evaluation, not the release of the model.

What changes

The debate concerns not just a model's release, but the distribution of its capabilities and the safeguards accompanying it.

Practical implication

Monitor defensive implications; do not confuse open-weight with intrinsic security.

Availability

Researchable reports; results from benchmarks and specific evaluation conditions.

Limits and uncertainty

Anthropic is also a commercial competitor. NIST evaluation provides a separate point of comparison; no ranking applies to all tasks.

Sept. 29 · NIST study on the 17th

Verified facts

• Space brings together files, Pages and project-related work; it replaces Library for eligible accounts, while Projects remains separate.
• Pages makes it possible to create and edit documents and interactive content with collaboration. The recap announces the collaborative slides for the following weeks, not as an immediate release.
• Meetings is in beta macOS for Pro and Business: notes and actions are recorded in Space, then audio is deleted.

What changes

To work with many, the shared object becomes the document and its sources, not just a copy of an assistant's answer.

Practical implication

Try a Page on a non-sensitive folder; check who can read and edit.

Availability

Space/Pages: Eligible Pro, Business and Enterprise offers. Enterprise share in preview opt-in; Meetings in beta macOS Pro/Business.

Limits and uncertainty

Do not confuse a shared page with sharing private conversations or account memory.

Verified facts

• First cohorts in San Francisco, New York and London.
• Participation on appointment of organizations.
• Twelve-week residence and practical assessment.

What changes

Production deployment and training are becoming a strategic focus.

Practical implication

Apply for eligibility only if a real business project is identified.

Availability

Initial US/UK cohorts; ask the account team about Belgian eligibility, which is not guaranteed.

Limits and uncertainty

10,000 engineers is a goal, not an already trained workforce.

Verified facts

• Claude Code goal: 50% of developers at the end of 2026 and a majority in 2027.
• The knowledge assistant, active since 2025, has already been adopted by more than 16,000 employees.

What changes

Agents are progressing in a large regulated bank.

Practical implication

Retain the importance of permissions and human control, not the only volumes.

Availability

Internal deployment Barclays; not a public offer Belgium/EEA.

Limits and uncertainty

The assistant and the usage volumes are existing results, not as many launches on 1 October.

Verified facts

• The current effective date is 1 December, postponed from 2 November.
• The advertised default limit is 4,000 Copilot Credits per user per month, which can be modified by an administrator.
• Belgium is explicitly one of the markets not initially available.

What changes

Agents need spending governance separate from licensing.

Practical implication

Do not transpose this configuration to a Belgian tenant; monitor the geographical extension and control the budgets.

Availability

Belgium explicitly excluded from initial deployment; several other EEA countries are also excluded. This is not a general activation in October.

Limits and uncertainty

Entry of 1 October updated on 14; read its revised effective date, not the first announcement.

Verified facts

• The described route returns to Safeway to pay.
• Expansion to other Albertsons brands is planned, not presented as already complete.

What changes

Assistants become entry points for trade.

Practical implication

Follow local conditions before recommending this route in Belgium.

Availability

US distribution context; no availability established in Belgium/EEA.

Limits and uncertainty

Do not present the payment as fully realized in ChatGPT: the body specifies a referral to Safeway.

Verified facts

• Google announces Gemini 4 Argon on September 30 and starts a progressive access for trusted cyber defenders via Fairwind.
• The announcement targets long-term professional tasks: development, finance, law and cybersecurity.
• The general availability of 13 October is the subject of a separate sheet; do not confuse this initial announcement with the subsequent deployment.

What changes

The October 1, 14 and 15 newsletters must not artificially become three launch dates.

Practical implication

Evaluate the API version currently available on a verifiable task with explicit budget.

Availability

As of September 30: limited progressive access, no general availability. See the October 13 update for API and application.

Limits and uncertainty

Retrospective addition: this announcement was missed in the pilot. Performance claims in the post are Google's and were not reproduced here.

Verified facts

• Dots is an always-on agent powered by GPT-6 Astra, with its own cloud computer, that can pursue a goal between conversations.
• The Pro rollout excludes the EEA, Switzerland and UK at launch; Enterprise access is a beta disabled by default.
• Usage temporarily does not count against the allocation during the first month for eligible plans; subsequent terms are not yet detailed in the notes reviewed.

What changes

For your projects, separate read-only preparation from changes, messages and publication. Preserve that design boundary in any autonomous assistant.

Practical implication

To be monitored, not to be integrated immediately: European Pro access closed at launch.

Availability

Belgium: no Pro access to launch. Do not extrapolate this restriction to professional offers without checking their conditions.

Limits and uncertainty

An agent available continuously is not a guarantee of correct result. Authorizations and review remain necessary.

Verified facts

• GPT-6.1 Sol succeeds GPT-6 Sol for agent coding, computer use and professional work; proximity to Astra is an assertion of OpenAI, not a test performed for this journal.
• For standard API requests with up to 272,000 input tokens: $2 per million input tokens, $0.10 for cache reads and $10 for output; above the threshold, long-context rates apply to the entire request.
• The announced window is 1,050,000 tokens. Tool calls use Responses; Chat Completions is supported without tool calls.
• Supported effort levels are low, medium, high, xhigh and max. EU data residency is supported, but Fast mode is not supported with that residency.

What changes

Compare on a bug and a feature in Syndiko or Melete: tests passed, diff quality, duration, human interventions and total cost. Keep the scope and reasoning-effort budget identical.

Practical implication

Do a limited comparative test; avoid sending a whole repository into context by reflex.

Availability

API available; Work/Codex gradually deployed starting with Pro. API subscription and billing are two distinct channels.

Limits and uncertainty

The maximum context is not free, and the price per token does not measure the cost of a complete task.

Verified facts

• Opus 5.5 is announced on September 22nd and Sonnet 5.5 on 28. Both are available on the Claude platform and the main clouds mentioned by Anthropic.
• API prices per million tokens: Opus 5.5 at $4 input / $20 output; Sonnet 5.5 at $2 / $10. Claude subscriptions have their own limits.
• Anthropic announces more than 30% additional generation speed for both; the announced savings per task are 40% for Opus and up to 30% for Sonnet, on its own tests.
• Five-hour limits increase on plans including Pro and Max with Opus 5.5; a limit reset can be saved and used later. Haiku 5.5 is announced for later, not available in these announcements.

What changes

Suggested starting point: Sonnet for a well-scoped implementation; Opus for architecture, difficult diagnosis and independent review. This is a hypothesis to measure, not a universal ranking.

Practical implication

Compare medium and high before putting max on all workers.

Availability

Available; increased limits announced for several subscriptions including Max. No status of your personal account has been inspected.

Limits and uncertainty

Efficiency figures are the vendor's. Le Signal AI is a practitioner's account, not proof that Opus wins on every task.

Sonnet update September 28; reminder of Opus launch on 22. Both dates are distinct.

Verified facts

• Can rewrite a prompt, intercept a tool and add an interface.
• Available in the CLI and desktop application.
• Mods are not sandboxed and have the same machine access as Claude Code.
• A sec-default mod protects certain controls in managed environments.

What changes

The agent becomes a much more powerful and privileged extension platform.

Practical implication

Install only trusted code and test managed controls.

Availability

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

Limits and uncertainty

Date of October 1st confirmed by the metadata and public index, not by the receipt of the newsletter of 3. The plugins portal is earlier: September 25th.

Verified facts

• The models are available on Workers AI and announced under Apache 2.0 license on Hugging Face.
• They are compatible with Jev API and have a vision encoder and 64k context window.
• The decision-making mechanism scores the choices of the scheme rather than generating an intermediate token reasoning by token.

What changes

A structured decision can avoid mobilising a large generalist model for each branch of workflow.

Practical implication

Test routing, calibration and referral rate to a human on a real set of tickets.

Availability

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

Limits and uncertainty

New Clef release on October 1, distinct from Jev/Ollama support already covered in September. Benchmark scores are Cloudflare's.

Verified facts

• MAI-Transcribe-2-Streaming handles 60 languages with continuous language detection and initial hypotheses in just over 100 ms, according to Microsoft.
• The advertised introductory price is $0.54 per audio hour until the end of the year.
• MAI-Voice-2.1 and Flash support 23 languages and 26 regional variants, with announced rates of $22 and $15 per million characters.
• The three models are accessible via Microsoft Foundry and MAI Playground; LiveKit is indicated as coming.

What changes

The partial transcript may allow the agent to prepare his answer before the end of the sentence.

Practical implication

Measure latency end-to-end, errors in French and consent before clone a voice.

Availability

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

Limits and uncertainty

Rankings and latencies cited come from Microsoft's post; check regional availability and EEA residency.

Verified facts

• The official website offers ESP32 and Linux SDKs and examples for Raspberry Pi, screens and sensors.
• The firmware and SDKs are under Apache 2.0 license and provided without warranty.
• The Muse Home Link offer is reserved for Muse subscribers in the United States; the application is not an order.

What changes

An open hardware layer can extend agents beyond the browser.

Practical implication

Use a dedicated development card without opening the sensitive home network.

Availability

Open code for experimentation; access to the Muse service from Belgium is not established. Home Link is explicitly US-only.

Limits and uncertainty

October 2 is attested by TechCrunch; the official site reviewed is undated. Opening the SDKs does not mean opening the Muse model.

Verified facts

• The configuration retains GB10, DGX OS and the NVIDIA AI stack.
• Two 64 GB units can be connected and managed via Sync Cluster Assistant.
• The partner availability announced is October 23, starting at $4,999.
• The Model Launcher Sync, including an OpenCode configuration, is announced by the end of the month.

What changes

A new local service option; the date of announcement should not be confused with the shipment.

Practical implication

Size memory and throughput for your actual model before considering a purchase.

Availability

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

Limits and uncertainty

Price announced in dollars; Belgian price, VAT and availability unverified reseller. Performance gains specific to NVIDIA tests.

Verified facts

• Extension of supported models, not initial launch of these models.

What changes

Users of this preview can test recent models.

Practical implication

Check Dreams limits before changing the model.

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.

Verified facts

• The agent can read accessible content, click, type, scroll and work across applications.
• Copilot requests permission before controlling an application; organization policies can disable the feature.
• The activation in the CLI goes through /computer on.

What changes

Software without APIs becomes accessible to agent workflows.

Practical implication

Test in an isolated profile, with dummy data and limited authorizations.

Availability

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

Limits and uncertainty

Preview: do not assimilate interface control and transactional reliability.

Verified facts

• Steps can be sequential or parallel and exchange structured results.
• Workflows can request validation, suspend and resume.
• Available on all Copilot plans; experimental features must be enabled in the CLI.

What changes

Coded orchestration makes the process more explicit than the free delegation alone.

Practical implication

Prototype a reproducible review or diagnostic pipeline with human control before modification.

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.

Verified facts

• Guardian can recover previous instructions and the context of handoffs.
• The pending messages resume after reconnection avoiding duplicates.

What changes

Better continuity reduces context errors.

Practical implication

Update and test resume, permissions and tasks in queue.

Availability

Open source version; function eligibility according to configuration and server.

Limits and uncertainty

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

Verified facts

• Applications remain accessible in Workspace settings > Apps.
• Existing permissions and access controls remain applicable.

What changes

The management of extensions is central.

Practical implication

Control authorized marketplaces and application permissions.

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.

Verified facts

• The initial availability of the add-in dates from 17 September, not October.
• The Microsoft 365 administrator must also authorize the add-in.

What changes

A default configuration change to take into account.

Practical implication

Check the Word status in the console and Microsoft 365 permissions.

Availability

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

Limits and uncertainty

Date of effect explicitly provided for in the September note; no proof of activation in a particular customer space.

Verified facts

• Codex Cloud runs tasks in isolated workspaces with reusable environments, repositories and dependencies; work continues with the computer closed.
• Codex CLI adds voice and an /agents view for tracking delegated tasks, plus improvements to resumptions and worktrees.
• Code Review and Security Cloud add review of changes and repository scans; Security Cloud requires repository configuration, applicable permissions and billing.

What changes

For your repositories: prepare a reproducible environment, prohibit automatic deployments, request a short diff and a verifiable test command. Do not enable paid scanning without an explicit budget.

Practical implication

Pilot an isolated cloud task; keep independent review and pre-merge tests.

Availability

Variable offers and configurations; check real access in your workspace before confiding a repository.

Limits and uncertainty

A completed task in the cloud is not a validated change. You still have to run the tests, read the diff and control the security.

Verified facts

• Decisions API focuses Luna on finite-answer questions supplied by the developer, from text or images; the recap describes it as a limited preview.
• Agents API gains computer use in a hosted browser. The changelog separately identifies GPT-6.1 Sol's multi-agent beta in Responses.
• These new features are distinct interfaces and orchestration capabilities, not three new models to compare in the same table.

What changes

To rank, prefer a structured output and a test set. For a long workflow, evaluate recovery, status and permissions. Add sub-agents only if parallelism brings a measurable value.

Practical implication

Watch Decisions availability; test the three components separately.

Availability

Decisions: limited preview; multi-agent: beta. Access and conditions of the hosted browser to check in the API.

Limits and uncertainty

No quality or cost measurements of these APIs were performed in this pilot.

Verified facts

• Ollama 0.35 adds the decision models and the /v1/systemone endpoint: a state and named questions produce typed answers.
• Use cases cited include triage, routing and moderation; Nimble 9B, Tev1 4B and Tev1 0.8B are offered for local execution.
• The 91 ms per decision announced for Nimble comes from an example Pac-Man on M5 Max, not from a latency guarantee on your VPS.

What changes

For Syndiko, test request triage on an annotated dataset before taking any action. Include an unknown class and a confidence threshold; never trigger a sensitive action from the score alone.

Practical implication

Build a read-only classification prototype; no model was installed in this pilot.

Availability

Ollama 0.35+; material requirements, memory and latency to be measured on the target machine.

Limits and uncertainty

Local does not mean without cost: equipment, maintenance and quality of ranking remain to be evaluated.

29 Sept. · e-mail 30

Verified facts

• Plugin extensions add interactive panels, side views and file viewers in ChatGPT.
• The support of the MCP Events proposal allows compatible applications to start automations during events.
• Participating integrations may consume the Plus/Pro allocation with permission and ceiling per application; connecting an app does not increase the quota.

What changes

For a SaaS product, an integration can bring an interface and workflows into ChatGPT. The real engineering work remains permissions, idempotency, auditing and duplicate-event handling.

Practical implication

Explore the documentation for read-only integration before business actions.

Availability

Extensions and events subject to permissions and compatibilitys; consumption of allowance in participating apps in limited preview.

Limits and uncertainty

MCP Events is described as a proposal. This announcement does not establish universal support in all MCP clients.

Verified facts

• First cohorts in San Francisco, New York and London.
• Participation on appointment of organizations.
• Twelve-week residence and practical assessment.

What changes

Production deployment and training are becoming a strategic focus.

Practical implication

Apply for eligibility only if a real business project is identified.

Availability

Initial US/UK cohorts; ask the account team about Belgian eligibility, which is not guaranteed.

Limits and uncertainty

10,000 engineers is a goal, not an already trained workforce.

Verified facts

• The current effective date is 1 December, postponed from 2 November.
• The advertised default limit is 4,000 Copilot Credits per user per month, which can be modified by an administrator.
• Belgium is explicitly one of the markets not initially available.

What changes

Agents need spending governance separate from licensing.

Practical implication

Do not transpose this configuration to a Belgian tenant; monitor the geographical extension and control the budgets.

Availability

Belgium explicitly excluded from initial deployment; several other EEA countries are also excluded. This is not a general activation in October.

Limits and uncertainty

Entry of 1 October updated on 14; read its revised effective date, not the first announcement.