OpenAI's BioBasicsBench tests agents on 100 raw wet-lab tasks, and GPT-6 Astra scores 66.7%
The benchmark uses gel images, colony plates, microscopy and instrument exports from real experiments, graded against expert rubrics.
The benchmark uses gel images, colony plates, microscopy and instrument exports from real experiments, graded against expert rubrics.
A Google Research framework turns a vague concept such as healthy food into a classifier through scoping, borderline images and automatic prompt refinement.
Collinear AI publishes a public set of 100 tasks in which an agent must find and fix a real weakness in an open-source repository without breaking it.
OpenAI says it is more than quadrupling the midpoint of its medium, high and critical reward ranges and adds quality, loyalty and campaign bonuses.
OpenAI says Plus, Pro and Business plans will draw ChatGPT and Codex use from a single allowance measured in tokens; Instant chat will not count against it.
The signal comes from an Internet-Draft published the same day with the policy token __usage_ai-training; OpenAI says it will implement it in the coming months and keep its own crawler rules.
The model announced in early October is now downloadable; Mistral pitches it at defensive security, with 82% on Artificial Analysis' Cyber Index and a licence that limits model-as-a-service use.
One open-weight checkpoint transcribes 40 languages, translates 31, labels up to eight speakers in hour-long audio, and takes plain-language prompts; Granary v2 data ships with it.
A wake-and-sleep cycle moves short-term context into newly unlocked MoE experts, which Google says retains knowledge up to 10M tokens without catastrophic forgetting.
NASK's AI Safety Research Center, Poland's national AI Safety Institute, will work with OpenAI on evaluating frontier systems and finding vulnerabilities.