Frontier access stops being a budget line
-
Nº LXVI
- Date
- 30 Jul 2026
- Issue
- 66
- Stories
- Seven
- Editor
- ARC
OpenAI hands academia the keys, while DeepMind reportedly closes the door on AlphaFold.
OpenAI opens models to academics
OpenAI opened frontier models to academic researchers free of charge, starting with 10,000 and expanding to 100,000 through 2027. Those selected get the company's most advanced models, including GPT-5.6 Sol Pro, and can invite four collaborators from their own institution, OpenAI told Axios ahead of the announcement. The program spans the sciences, mathematics, and engineering. Frontier-model access has been the budget line separating well-funded groups from everyone else; for the next 18 months that gap narrows across academic biology.
AlphaFold team reportedly disbanded
Google DeepMind disbanded its AlphaFold team as strategy consolidates around Gemini, according to a market-chatter report circulating on Hacker News. DeepMind has not confirmed it. If the report holds, the structure-prediction franchise that made AI credible to biologists gets absorbed into a general-purpose model, and who maintains dedicated biological models becomes an open question for the field.
Also discussed on X.
Synthetic genomes anchor OpenAI paper
A GPU-native genome engine sits at the center of a new paper co-authored by OpenAI and Minos AI on scientific computing in the age of agentic AI. HelixForge, Minos AI's engine for generating synthetic genomes, is featured as the study's worked case. Using genome-scale generation as the test system puts biology at the center of how agent-driven scientific computing gets evaluated.
Ensembles beat static docking
Conformational ensembles beat static structures for protein-ligand pose prediction, substantially outperforming them in a category-stratified benchmark from the Mavchen 1 platform. Pose-prediction accuracy claims now carry an implicit question about which structural regime produced the number.
Generalist models rank variants
General-purpose language models tested on protein variant ranking in a new benchmark, PG-LLM, pitting models with no protein-specific training against the variant-effect task. Gives the field a reference point for whether specialist protein language models still hold their edge.
Agents try open-ended research
Agents ran open-ended research in two case studies, producing early evidence on where autonomy holds up and where it falls apart. Grounds the autonomous-discovery debate in observed runs rather than benchmark scores.
EvoPINN searches algorithm space
EvoPINN evolves executable algorithms for physics-informed neural networks, models constrained by known governing equations. Agentic search over algorithm code pushes toward the simulation stack biology leans on for tissue mechanics and pharmacokinetics.
Reply with your discoveries. A human reads them. Forward freely.
|