Coverage: 28 September–4 October 2026. Reporting cutoff: 21:53 PKT on 4 October.
OpenAI released GPT-6.1 Sol, Anthropic launched Claude Sonnet 5.5, and Google announced a restricted rollout of Gemini 4 Argon. The key difference for anyone choosing an AI tool this week is access: the first two can enter everyday workflows, while Argon remains limited to selected testers.
This AIX Radar weekly AI news roundup covers 28 September–4 October 2026. Alongside the model launches, we examine reusable agent workflows, an open-weight German–English model, voice APIs and the security controls needed when AI can take actions.
This week’s AI releases: what can you use?
| Announcement | Access at announcement | What to check |
|---|---|---|
| GPT-6.1 Sol | API, Codex and eligible ChatGPT Work plans | Work access differs from regular Chat. |
| Claude Sonnet 5.5 | Claude Platform and major cloud platforms | Task-cost savings are vendor-reported. |
| Gemini 4 Argon | Selected cyber defenders and testers | General access is still planned. |
| Kolibri | Downloadable weights, Apache 2.0 | Evaluate hosting requirements and context length. |
| Gemini Skills | Chat rollout; Workspace to follow | Plan for the transition from Gems. |
Also covered: OpenAI’s agent products, Meta’s business tools, Microsoft voice APIs, Supabase–Turso, agent security and AI research.
29 September 2026
GPT-6.1 Sol launches at $2 input and $10 output per million tokens
OpenAI released GPT-6.1 Sol in its API, Codex and ChatGPT Work for eligible paid plans. The launch announcement says it is not yet available in regular Chat. Its standard API rates for prompts up to 272,000 input tokens are $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million. Those standard input and output rates are one-fifth of GPT-6 Astra’s.
OpenAI describes performance approaching Astra on several evaluations. That is a vendor-reported assessment, not evidence that the models are interchangeable for every task. The API release also adds beta support for delegating work to multiple agents.
For repeated coding or document work, compare cost per correctly finished task, including retries and review. A cheaper token rate does not settle that comparison.
OpenAI: Introducing GPT-6.1 Sol · OpenAI API changelog, 29 September
28 September 2026
Claude Sonnet 5.5: faster output, unchanged token prices
Anthropic introduced Sonnet 5.5 for tasks such as bug fixes, documents and spreadsheets. It is available through the Claude Platform and major cloud platforms, including AWS, Google Cloud and Microsoft Azure.
The company reports output generation more than 30% faster than Sonnet 5 and task costs up to 30% lower in its testing. The distinction matters: published input and output rates remain $2 and $10 per million tokens. Anthropic attributes the task-cost reduction to using fewer tokens, rather than cutting those rates.
For a migration test, rerun the same bug fix or document task and record completion time, errors and total token use. That will show whether the claimed efficiency reaches your workflow.
30 September 2026
Gemini 4 Argon: restricted access and introductory API pricing
Google announced Gemini 4 Argon for demanding software engineering, enterprise work and cybersecurity defense. Initial access is limited to trusted cyber defenders through its Fairwind Program and selected testers. A broader rollout is planned, beginning with paid API customers and Google AI Ultra subscribers.
Google lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 respectively after the introductory period. It also describes a one-million-token output limit for extended tasks. That output limit should not be confused with the size of an input context window.
Argon is a significant announcement, but readers should not assume they can access it today. Availability, final pricing and independent evaluations will determine how it compares in practical use.
3 October 2026
Kolibri: Apache 2.0 weights for German and English AI applications
Aleph Alpha released Kolibri with downloadable weights under Apache 2.0. The model supports reasoning and tool calling, with 78 billion total parameters and approximately 3.46 billion active per token. Its focus is German and English applications.
The long-context specification needs care. The official model card lists a native context of 262,144 tokens and says the company validated extensions up to 1,048,576 tokens. It recommends staying within the native length for latency-sensitive deployments and complex tasks.
This gives teams another option for running a model on infrastructure they control. Evaluate it on your own documents and language requirements; open weights and company compliance claims do not establish that a particular deployment meets every regulatory obligation.
Aleph Alpha: Kolibri release, 3 October · Official Kolibri model card
29 September 2026
OpenAI DevDay: Dots, ChatGPT Space and expanded Codex access
OpenAI’s DevDay announcements included Dots, agents intended to handle ongoing responsibilities; ChatGPT Space and Pages for shared work; and expanded cloud access for Codex. Dots launched on eligible Pro and Business Premium plans, with an administrator-enabled beta for certain organizational plans.
The developer announcements included a limited-preview Decisions API, computer use in the Agents API, and new plugin capabilities. Availability differs between products and plans, so the event should not be read as a universal rollout of every feature.
The direction is clear: AI products increasingly compete on what happens between conversations. For a small business, the important questions are what an agent can access, how its work is reviewed and whether it reliably remembers an ongoing assignment.
30 September 2026
Gemini Skills begin replacing Gems with reusable workflows
Google began rolling out Skills in Gemini chat: saved instructions that can be invoked with a slash command, combined with other skills and supplemented with reference files. The announcement says Workspace availability will follow in the coming weeks.
Skills will replace Gems in stages. Google says personal-account Gems will begin retiring in November, with later transitions for business and education accounts. Existing Gems are scheduled to migrate automatically; the separate Gems by Google Labs experiment is an exception.
Writers, marketers and small teams should document the instructions they reuse most often—such as a brief format or fact-checking process. A saved workflow is valuable only if its instructions are specific enough to produce a result you can assess.
28–29 September 2026
Meta introduces its Enterprise Platform and Muse for Small Business
Meta announced its Enterprise Platform on 28 September, naming Chirantan “CJ” Desai to lead it. The stated scope includes the Muse agent, Meta Business Agent, Muse API and Muse Code.
A day later, Meta introduced Muse for Small Business, adding skills and connectors for business tasks. The announcement builds on Muse’s availability in the US and Canada and describes assistance such as identifying emails that need attention and preparing drafts. It does not establish that every region or business application is supported.
Distribution is the story here. Many small businesses already use Meta’s services, so an agent attached to existing business activity may be easier to encounter than a separate enterprise tool. Useful follow-up coverage should test the actual connectors, approval controls and completed work.
Meta: Launching Meta Enterprise Platform · Meta: Muse for Small Business
1 October 2026
Microsoft launches MAI streaming transcription and voice APIs
Microsoft introduced MAI-Transcribe-2-Streaming alongside MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The transcription model supports 60 languages and is offered at an introductory $0.54 per hour of audio through the end of the year.
Microsoft says the streaming model starts producing partial transcripts just over 100 milliseconds after receiving audio. Its new voice models support 23 languages and 26 locales. Listed text-to-speech prices are $22 per million characters for Voice-2.1 and $15 for Flash.
These are relevant building blocks for live captions, transcription products and voice assistants. Published latency is not a complete application test: recording conditions, network delay, corrections and the handling of names still need evaluation.
2 October 2026
Supabase’s Turso acquisition targets databases for AI agents
Supabase and Turso announced that Turso is being acquired by Supabase. Their stated goal is lightweight databases that agents can create on demand, with a path into Postgres as applications grow.
Both companies say Turso will continue operating. Supabase remains committed to Postgres, while Turso continues its SQLite work. Turso says deeper integration will follow in the coming months, so the announcement should not be read as a completed rollout of every proposed integration.
AI coding tools make it easier to generate applications; their databases still need isolation, backups and a manageable operating cost. For existing users, this announcement is a reason to follow the integration roadmap and portability options, rather than immediately migrate a working application.
Supabase: Acquisition announcement · Turso: What the acquisition means for users
28 September 2026
NVIDIA OpenShell and Sentry add controls around AI agents
NVIDIA announced the Open Agent Safety Platform, combining OpenShell runtime software with its Sentry reference system design. OpenShell applies policies around an agent’s execution. Sentry runs on BlueField-4 hardware and monitors behavior from outside the agent’s own environment.
NVIDIA says Sentry can quarantine agents that breach their boundaries within milliseconds. That is the company’s stated capability, not a guarantee that every deployment or attack is contained. OpenShell is broadly available; the announcement describes Sentry as a reference system design.
This addresses an important design question: who stops an agent when the agent itself is the source of the problem? Independent enforcement is worth watching as tools gain permission to edit files, use business applications and run for longer periods.
1 October 2026
Microsoft’s security report highlights agent permissions and identity
Microsoft’s 2026 Digital Defense Report describes attackers using AI within reconnaissance, social engineering, malware development and other existing workflows. Its accompanying analysis also emphasizes the systems surrounding legitimate agents: identities, tools, data access and the ability to revoke permissions.
An AI security review should not stop at the model. For a website owner or SaaS team, it should also ask which account the agent uses, what that account can change and how access can be withdrawn. Faster automation increases the value of getting those basics right.
29 September 2026
Microsoft Quine enters limited research access for biology
Microsoft Research introduced Quine, a research system combining a model of biological processes with scientific tools, literature and experimental workflows. Initial access is limited to a fellows program and selected research collaborations.
The research announcement describes using the system to prioritize candidate compounds for laboratory testing. Those experiments are early research evidence, not proof of an approved treatment or a replacement for clinical validation.
Quine illustrates a different use of AI from consumer chat: helping researchers choose what to test next. The meaningful result will be the quality and reproducibility of the experiments it helps produce, rather than a fluent explanation of biology.
AI research, industry and policy briefs
2 October: Meta shares mathematics research developed with Muse Spark
Meta published six papers from collaborations between mathematicians and Muse Spark, saying five address previously open research questions. Its account describes researchers guiding the work and a second group reviewing it; it also acknowledges independent work on some of the same problems.
This is a research announcement about expert–AI collaboration. The practical point is the documented cycle of proposing, checking and correcting arguments. It should not be treated as independent confirmation that the model can autonomously solve arbitrary research problems.
Meta AI Research: Papers, collaborators and review process
28 September: Mistral opens an industrial AI hub in Munich
Mistral announced a Munich hub for industrial and physics AI, with research teams and applied engineers supporting enterprise partners. It identified work with BMW and Siemens Energy. The development adds a European industrial perspective to a week dominated by general-purpose agents.
Mistral: Munich industrial AI hub
29 September: US policy signals a preference for industry-led oversight
Associated Press reported that the White House and major technology leaders signed a voluntary AI accord involving internal and external reviews. This is a policy development to follow, not evidence that the industry’s safety questions have been resolved.
Associated Press: Voluntary AI accord
Separately, Executive Order 14434 directs changes to terminology used within the US executive branch, favoring “Super Intelligence.” The terminology is the administration’s policy language; it does not by itself establish a scientific finding about model capabilities.
What this week changes for AI tool buyers and developers
For developers: compare models on a small set of real tasks with measurable pass conditions. Include review time and failed attempts in the cost. Keep restricted previews separate from tools available in your account.
For creators and small businesses: start with one workflow that repeats frequently, such as preparing a content brief or reviewing incoming requests. Decide which steps can be automatic and which need your judgment before connecting more applications.
For teams adopting agents: document the data and actions each agent can access. Check logs and revocation controls as carefully as output quality. This week’s security announcements are a reason to ask sharper purchasing questions, not a substitute for testing a particular product.
AIX Radar will follow the major stories individually as useful evidence emerges, alongside this weekly briefing. Find related guides in our AI coding coverage and future briefings in AI News.
Sources and reporting cutoff
Updated 21:53 PKT on 4 October: added the 2 October Supabase–Turso announcement, 3 October Kolibri release and 2 October Meta research papers after verifying their original sources. These events occurred earlier in the week.
We checked original announcements and documentation across models, business tools, developer products, security and AI research. The policy accord is attributed to Associated Press reporting. Performance claims are attributed to the companies. The practical assessments are AIX Radar’s editorial analysis; they do not represent hands-on product testing.
The coverage window is seven calendar dates in Asia/Karachi, with 4 October included only up to the reporting cutoff. Dates beside stories follow their original announcements, not the dates of later reposts. This is a selection of significant developments, not a claim to catalogue every AI announcement. Unconfirmed leaks and stories whose underlying event fell outside the window were excluded.
