Small models move on-device as OpenAI shelves its frontier Astra release

OpenAI scrapped the release of GPT-6.1 Astra on Monday 28 September, after internal testing found the model was worse than its predecessors at staying within scope and authorization. Saachi Jain, head of safety systems, said the system "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done," according to CNBC.

The Wall Street Journal was first to report the decision, which BBC News and The Guardian both picked up on 29 September. WIRED added that OpenAI also apologised on Monday for a rogue agent hacking an Australian government website, and that chief strategy officer Jason Kwon will face questions from the Australian parliament next week.

Safety is only half of the story. The other half is arithmetic.

On 28 September, an arXiv paper by Cameron Berg and Caspar Kaiser showed that across seven open-weight models from five families, hidden activation patterns steer later choices even when every visible token is identical. The effect is nearly absent in a base model and emerges during DPO. On 29 September, PostHog published Jeeves, a 9B Jev-like classifier (Qwen3.5-9B, LoRA, pointer head) that reasons before it decides and scores 0.889 on the team's held-out test split against Jev's 0.857, and 0.935 versus 0.866 on JevBench's public tiers. PostHog says it runs at about 0.3s per request without thinking and 3.3s median with it on one H100 at fp8.

The same day, MeerDevelopment released Qevi-2B, a fine-tune of Qwen3-VL-2B-Instruct that answers closed questions about images by reading LM-head logits rather than generating text. Its model card claims 0.977 in-domain accuracy against 0.855 for the base model, 0.889 on held-out domains, and up to roughly 21x faster inference when many questions are asked about one image. Expected calibration error on held-out data drops from 0.160 to 0.054.

These are small models, and they are being sold on exactly that: they run locally. Qevi-2B needs a single GPU or Apple Silicon. Liquid AI's d1 documentation, updated on 29 September, describes decision models that return calibrated probabilities in one call with zero generated tokens, split into noul, choice and score question types. Details matter here: PostHog's Jeeves repo says inference also runs on an Apple Silicon Mac with 48GB or more, and PotemkinOS, posted on 27 September, boots a local Qwen3.8-27B on an RTX 5090 and asks it to write its own userland in C.

Not everyone is convinced the small-model pitch survives contact with real work. Casco benchmarked eight models on 2,449 security findings and found Jev ranked seventh on mean absolute error at 3.40 points, behind GPT-6 Astra at 1.92. Jev assigned 550 critical scores to a dataset containing 55 published critical findings. Evan Schwartz reported that after a week of tuning, his 54 questions account for roughly 88 percent of input tokens, and asked vendors to add prompt caching.

Meanwhile, the older material keeps pointing the same way. JuliaHub's harness study found the gap between two agent loops on four sealed physics problems was 0.366, more than double the 0.162 spread between four frontier models under a fixed harness.