🤖 GPT-5.6 Preview ships under White House access controls 🤖
White House restricts access to OpenAI's most capable model family yet.

OpenAI released GPT-5.6 Preview, a family of three models — Sol, Terra, and Luna — with Sol reportedly outperforming Anthropic's Mythos 5 on Terminal-Bench 2.1 using a third of the output tokens. In an unusual move, the White House asked OpenAI to limit initial access to around 20 vetted government partners, citing safety concerns after evaluator METR flagged higher cheating rates on evaluations. CEO Sam Altman said a general release would follow within a couple of weeks.

💰 Claude Sonnet 5 ships at $2 per million tokens 💰
Anthropic's newest Sonnet-tier model targets agentic workloads with aggressive pricing.

Anthropic launched Claude Sonnet 5, which the company says approaches Opus 4.8-level capabilities while beating Sonnet 4.6 across planning, tool use, and coding. The model can operate a browser or terminal autonomously for extended multi-step jobs. Introductory API pricing lands at $2/$10 per million input/output tokens through August 31, directly pressuring OpenAI and Google on agentic pricing.

🧬 Anthropic ships Claude Science, enters drug discovery 🧬
A new research platform integrates with lab infrastructure and targets neglected diseases.

Anthropic shipped Claude Science, a research platform with over 60 specialized scientific tools that connects to existing lab infrastructure and automates complex analyses with full audit trails. Alongside the platform, Anthropic is launching its own drug discovery program targeting "neglected" diseases historically skipped by pharmaceutical giants. The move positions Anthropic as a direct player in healthcare R&D rather than just a model provider.

INTERESTING TO KNOW

🧠 Meta's Brain2Qwerty v2 hits 78% word accuracy 🧠

Meta released version 2 of its Brain2Qwerty non-invasive brain-computer interface, which now reads full words and their meanings rather than individual characters. Nine volunteers generated nearly 22,000 sentences of training data, pushing accuracy to 78% for the top volunteer — up sharply from the 8% highs of earlier non-invasive approaches. Meta open-sourced the code for both versions.

🍉 Meta's Watermelon Catches Up to GPT-5.5 🍉

Meta's superintelligence chief Alexandr Wang revealed that its upcoming model, codenamed Watermelon, has matched OpenAI's GPT-5.5 on closely watched benchmarks. The model is still in training and reportedly uses an order of magnitude more compute than Muse Spark, with no release timeline announced yet.