GPT-6 Astra outscores every specialist at antibody prediction
The rule that AI-for-science requires dedicated models just broke. OpenAI's GPT-6 Astra topped a blind benchmark for predicting whether antibodies can become drugs — beating purpose-built systems without any domain-specific training.
GPT-6 Astra scored 37.98 on the DDD Benchmark's antibody developability module, outperforming every specialized model in a blind evaluation built by Insilico Medicine. The benchmark integrates six sets of antibody experimental data and tests a question that has burned through billions of drug-development dollars: not whether an antibody can bind to a target, but whether it will aggregate, remain stable, and function as a real drug in the body. Andrew Aiginin, the scientist who published the result, emphasized that GPT-6 achieved this without external tools or fine-tuning on antibody data — it drew on general knowledge alone. OpenAI president Greg Brockman reposted the announcement on X, lending institutional weight to the finding. The result comes from the same DDD Benchmark system Insilico Medicine detailed in a 21-page paper on arXiv, which also showed that chemical-aware LLM training can beat traditional retrosynthesis models. If a general model can internalize enough physics and chemistry to predict druggability, the pharmaceutical industry's expensive pipeline of custom AI tools faces a serious question about whether specialization is still worth the cost.
Raycast 2.0 ships with a full redesign, a custom file indexer, and AI that loads your agent context automatically. The macOS launcher's biggest overhaul since its 2023 debut adds built-in dictation (hold a hotkey, text appears system-wide), a unified Quick AI and AI Chat composer that shares context across conversations, and automatic loading of Skills installed on your Mac — meaning any agent setup you've built elsewhere is available inside Raycast's chat. Files, folders, and contacts now surface alongside apps in Root Search via a custom indexer Raycast built to replace Spotlight, which the team says never delivered the speed or consistency they needed. During the beta, OpenAI's GPT-5.4 mini is free for everyone, along with global dictation and several Pro features. The update targets developers who already live in the keyboard-shortcut workflow but want AI woven into the OS layer rather than bolted on as a separate app.
What to watch: whether Insilico publishes a full leaderboard with scores for competing models — the current coverage cites GPT-6's number without showing the runner-ups, and that gap matters.
Should drug companies trust a general-purpose LLM over a model trained specifically on their molecular data? Tell us in the comments.
