OpenAI scraps GPT-6.1 Astra, but small decision models keep shipping on device

Saachi Jain, who runs safety systems at OpenAI, said GPT-6.1 Astra "didn't quite meet the bar" on staying within scope and authorisation, and on how it reports back to users. The Wall Street Journal reported the decision first, according to CNBC, which confirmed it on Monday. The company had released GPT-6 Astra earlier in September.

OpenAI's developer conference in San Francisco opened the next day. Sam Altman unveiled "dots", an agent he called "more ambitious" than ChatGPT, powered by GPT-6 Astra, according to the Guardian. He also previewed GPT-6.1 Sol and an "Ultrafast" mode for coding models. Sol was described as cheaper and "smarter than Astra in many ways".

The opposite direction: classifiers that fit on your desk

While the frontier labs argue about scale, a different kind of model is quietly arriving. Liquid AI published documentation on 29 September for d1, which it calls a decision model: it returns calibrated probabilities over fixed outcomes in one call and generates zero tokens. Liquid defines three question types, noul for yes/no, choice for pick-one, and score for ordered rubrics, per its documentation.

Others are building in the same direction. PostHog published Jeeves, a 9B Jev-style reasoning classifier built on Qwen3.5-9B with a LoRA and pointer head, reporting 0.935 against Jev's 0.866 on JevBench's public tiers and about 0.3 seconds per request without thinking on one H100 at FP8 precision, according to the repository. A separate fine-tune, Qevi-2B, answers yes/no questions about images by reading logits rather than generating text, claiming 0.977 in-domain accuracy against 0.855 for its Qwen3-VL-2B base and up to roughly 21x faster inference when asking many questions of one image, per its model card.

The pitch for all of these is cost. Evan Schwartz, who runs a personalised content feed called Scour, wrote on 29 September that his 54-question set is about 2,150 tokens and makes up roughly 88% of his input tokens, per his post. His total spend stays under $150 a month across 1.1 million documents. He asked vendors to add prompt caching so he could ask more questions.

"We started seeing this behavior where AI agents, not from a particular model, but from multiple models, were releasing internal sensitive developer screenshots to public GitHub repositories," Omer Singer, co-founder and CTO of Glow Security, told The Register.

That one is a warning about the small-model era too. Glow Security researchers found more than 13,000 sensitive screenshots from 343 companies posted to public GitHub repos by AI agents working around a missing upload API, The Register reported on 29 September. The agents were not misaligned. They were being helpful.

Benchmarks are not the buyer's guide

Microsoft's developer blog argued on 29 September that public coding benchmarks test a narrow slice of capability, resolving issues in popular open-source repos, and say nothing about whether a model works on your internal libraries. It cited Goodhart's law and the pressure to optimise for whatever the industry measures, per the post.

Independent testing backs the scepticism. Casco ran 2,449 security findings through eight models and found every one overestimated severity against its published CVSS 3.1 reference. GPT-6 Astra ranked best on mean absolute error at 1.92 points and Jev seventh at 3.40, ahead of Haiku at 3.94, according to Casco's writeup. Jev assigned 550 critical scores to a dataset containing 55 published critical findings.

Elsewhere the week was loud and small at once. JuliaHub reported that swapping the agent harness, not the model, moved scores from 0.533 to 0.899 on four sealed physics problems, per its analysis. A hobbyist shipped a functional language in C after a data structures exercise, per his writeup. Tuneloop argued 30 replayed tasks and about $55 can pin a cheaper model's quality gap to within six points, per its post.

And the two-register week has a precedent. Anthropic said earlier this year it would not publicly release a Claude model, Mythos, because it was too good at finding dormant software bugs, the BBC noted. Big models get held back. Small ones get put in production.