OpenAI makes its agent infrastructure available to developers
The Agents API entered public beta on September 10. Developers can use the software coordinating Codex rather than build it themselves.
It manages long sessions, saves relevant context as conversations grow, finds tools when needed, and delegates parallel tasks to separate agents. You supply the task and tools.
OpenAI hosts the coordinating software. The agent's code and files can run in an OpenAI-managed sandbox, an isolated computing environment, or on your own infrastructure or a partner's.
There is no additional Agents API fee, but model tokens, tools and computing resources still cost money. Public beta also means the service is still changing.
Try it before maintaining your own agent framework. It does not remove the need for access controls, spending limits or human approval of consequential actions.
Anthropic's CEO calls for slowing the AI race
Dario Amodei wants frontier AI development to slow down. He argues that AI helping build its successors could outrun labs' ability to understand and control them.
Anthropic commits to embedded external evaluators with employee-like access and publication rights free from its editorial control, subject to confidentiality and security exceptions. Anthropic says it will invite the team soon.
Amodei also proposes coordinated safety standards and limits across labs, followed by international agreements. These remain proposals; Anthropic has not announced a blanket training pause.
The call follows Anthropic's disclosure of a fourth cyber-evaluation incident. An older Claude checkpoint accessed a real external system during a misconfigured test with production cyber safeguards disabled. METR will independently investigate the incidents.
Sam Altman backed the proposal and said OpenAI would also give independent evaluators employee-like access. Elon Musk and Demis Hassabis expressed support too. These are public endorsements and commitments, not an agreed industry-wide pause.
The next test is implementation: what access reviewers actually receive, what they can publish, and whether shared safety standards follow.
AI tackles mathematics beyond the leaderboard
OpenAI published a claimed solution to the Navier-Stokes Millennium Prize problem: whether equations describing fluid motion can break down despite smooth starting conditions and a smooth applied force.
An unreleased model coordinated roughly 10,000 agents, reaching a result after 88 hours. Astra then helped convert it into a proof checked by Lean, software that verifies mathematical reasoning.
That is substantial evidence, but human review still matters: experts must confirm that the formal statement matches the intended problem. Credit and the relationship to another team's work are also disputed.
Separately, Anthropic formalized Fermat's Last Theorem in 11 days. Claude converted an existing human proof into a computer-checkable one; it did not discover the theorem's original proof.
For researchers, the useful development is the combination of AI-generated work and machine-checkable verification. Neither announcement means a normal chatbot can reliably solve arbitrary research problems.
Quick hits 🗞️
-
ChatGPT Images 2.5 is rolling out across all tiers. It adds sketch-guided generation, comments for targeted edits and improved consistency across revisions. OpenAI claims up to 50% lower generation latency. Flare and Sunburst versions are available through the API. Source
-
GPT-Live-1 reaches developers. The voice model can listen and speak simultaneously, handle interruptions and delegate work to other models. Pricing starts at $0.05 per minute for the voice layer; backend reasoning and tools add to the bill. Source
-
DeepSeek is changing existing API routes. V4.1-Flash adds native vision and public weights. From September 14 at 04:00 UTC, even deepseek-v4-pro requests switch to Flash until V4.1-Pro arrives. If your application uses that name, test the replacement now. Source
-
Meta launched Muse in the US. The personal agent works in the background through its app or WhatsApp, using a dedicated cloud computer. Meta says sensitive actions require permission. Free access is available; stronger encryption is promised later this year. Source
-
Nvidia confirmed the Hugging Face deal. Its filing specifies $11.9 billion for shareholders plus up to $1 billion in employee retention awards. Closing is expected in the first half of 2027, subject to approvals, with a commitment to support other chipmakers. Source
See you next week!
If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.
