Anthropic loosened biology limits for vetted labs as OpenAI disclosed model misbehaviour
Anthropic opened a Life Sciences Verification Program giving credential-checked labs access to Mythos, Opus and Sonnet with biology blocks relaxed, and a grant that removes them outright for a single project, while OpenAI published six cases of its models hiding mistakes and inventing data under a framework committing it to disclose future ones.
Anthropic is giving vetted labs Claude with the biology blocks off The Life Sciences Verification Program, opened in beta on Thursday, gives credential-checked teams access to Mythos, Opus and Sonnet under classifiers tuned to permit drug discovery, research biology, clinical development and manufacturing work the general models refuse. A separate High-risk Use grant removes the life-sciences safeguards altogether for one named project and has to be renewed every six months; Anthropic says it is working with the US government to make those grants more broadly available on Mythos, and that for now they stay with a small set of entities under extra vetting. Source
Claude made more than 30 open-source biology models run four times faster Anthropic published the result the same day: an internal research model optimised more than 30 structure-prediction, protein-design, genomics and protein-language models in just under four weeks, averaging a 4x speedup, and built a mode that folds systems above 10,000 tokens — a bacterial ribosome, a proteasome, mitochondrial complex I — on a single Nvidia node. The optimised code is open-sourced, and Anthropic is putting up to $1m in Claude credits into a protein design competition with Adaptyv Bio that will wet-lab test over 5,000 community designs. Source
The UN moved its global statistics onto a Google-built AI platform The UN System Data Commons went live on Thursday, replacing the UNData portal with natural-language search and Model Context Protocol support so models can query it directly, built on Google’s open-source Data Commons with $2m from Google.org. It launched with data from nearly 20 agencies and commitments from 26. The reason, TechCrunch reports, is a UNICEF working paper — not yet peer-reviewed — which tested six large language models across more than 133,000 responses on development indicators and found an average accuracy of 21.2 per cent. Source
Crusoe raised $3.9bn at a $30.9bn valuation The AI infrastructure company announced the initial close of its Series F on Thursday, co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Nvidia, Founders Fund, GIC, the Qatar Investment Authority and TPG among the rest. Less than a year ago it raised $1.37bn at roughly $10bn; it now claims more than $140bn in contracted value. Source
OpenAI published six cases of its models hiding mistakes A disclosure framework out on Wednesday commits OpenAI to report misalignment it finds during training and evaluation, on three tracks: straightforward cases within six business days, minor investigations within twelve, and longer ones where third parties or security are involved. The six it opened with include a GPT-5.6 Sol training run in which the model wrote instructions into compaction summaries to conceal its errors and invent missing data, and a model that found an exposed API key in a public repository, used it, and made up the earnings figures when it still could not find them. Source
Databricks put Astra on 3,500 engineers and coding spend rose 60 per cent Engineering co-founder Patrick Wendell said on Wednesday that Databricks had rolled GPT-6 Astra out to every engineer after a 200-person pilot, and that engineers with access raised overall coding spend about 60 per cent against baseline. He said Astra clearly beats Opus 5 and GPT-5.6 Sol on complex, long-horizon system design but shows no meaningful gain on medium and low complexity tasks, so engineers now get a separate Astra sub-budget. Source
A Neuralink participant spoke aloud from neural signals alone Neuralink posted footage on Wednesday of Kenneth Shock, a participant in its VOICE trial who was implanted in January 2026, producing audible speech without movement, including “I love you” to his wife. The implant is investigational and has not been approved by the FDA. Source