OpenAI's GPT-6.2 API can drop reasoning when history is edited
The Responses API also drops temperature and top_p, and OpenAI's new Astra and Sol prompt templates target terse answers and needless questions.
Models, from The Decoder
Today's stories share a preference for fewer moving parts and more checking. OpenAI trims API controls, Bytedance and Nvidia researchers argue that three training phases and shorter contexts often suffice, and Google adds policy and anomaly checks around agents. In Lean, a Claude-assisted proof earned certification where an all-human draft drew problems.
7 stories, about 9 minutes to read them all.
The Responses API also drops temperature and top_p, and OpenAI's new Astra and Sol prompt templates target terse answers and needless questions.
Models, from The Decoder
The open model loops 12 of its 24 layers three times per token, and a GitHub project squeezes it to 5.38 GB for Apple silicon.
Open source, from The Decoder
A paper with several US universities says stacked training pipelines hide which step helps; The Decoder calls its open model competitive.
Open source, from The Decoder
Semantic governance policies, in Preview, use an LLM judge to test each proposed agent tool call against written rules and the user's request.
Agents, from Google for Developers
The Private Preview tool reads OpenTelemetry traces and logs sent to a Google Cloud project, so it needs no access to the agent's runtime.
Safety and security, from Google for Developers
Periodic summaries and targeted lookups worked about as well in many tests, The Decoder reports. The full study details sit behind a paywall.
Research, from The Decoder
A separate all-human Stanford draft has drawn reported problems; its announcer, Ravi Vakil, makes no claim it is correct.
Research, from The Decoder