OpenAI takes GPT-6 Astra into law firms, Anthropic builds its own wet lab, and 3 more AI moves

In brief

• Legal AI: OpenAI released Astra for Law, a version of GPT-6 Astra wired to a legal search index and sold to law firms under a confidentiality program.
• Capital: Emulate, a British world-models company founded last month by three former Google DeepMind researchers, is reported to be closing a seed round of up to $700 million.
• Life sciences: Anthropic confirmed it has built a wet lab in the San Francisco Bay Area and wants Claude to direct experiments.
• Power: Emerald AI, Google, Nvidia and Anthropic launched an alliance to make demand response standard for data centers, with a target of 100 gigawatts of new grid connections.
• Regulation: an NPR investigation found Google helped draft chatbot-safety bills in at least ten states, with exemptions that would cover its own products.

Five developments

1. OpenAI sells a legal version of GPT-6 Astra

• OpenAI released Astra for Law on Thursday, a configuration of its newest model, GPT-6 Astra, for law firms and legal software companies. The product pairs the model with a search index covering U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, alongside built-in instructions for legal analysis and writing 1.
• On 200 questions from the private validation set of Vals AI's Legal Research Bench, OpenAI says Astra for Law passed the benchmark's overall correctness check on 54.0% of questions at the highest reasoning effort, against 38.7% for GPT-6 Astra working from web search alone. On case-law questions the legal configuration found 24% more reference cases and retrieved up to 54% more relevant passages from the correct court opinions. It appears in the model picker as "GPT-6 Astra Law" and will reach the API as gpt-6-astra-law 12.
• Access runs through a Trusted Access Program that OpenAI says includes zero data retention on its API and keeps law firms' ChatGPT Enterprise usage outside human review by default. Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins and Wachtell Lipton worked on testing and development, Harvey and Legora can build on the model, and 26 partner plugins connect ChatGPT to tools including Relativity, Clio and Thomson Reuters' HighQ. The competitive ground is crowded — Google expanded its Gemini offerings for legal professionals in August and Anthropic has been adding lawyer tools since January — and the two things OpenAI is selling here are the index and the confidentiality terms 12.

2. A month-old British world-models company is reported to be raising $700 million

• Emulate, founded in August by three researchers from Google DeepMind's world-model team, is in advanced talks to raise up to $700 million in a seed round, according to reports in the Financial Times and Bloomberg. The two outlets gave different valuations: the FT reported $3.7 billion before the new money, while Bloomberg said the company was seeking $3 billion. The terms have yet to be agreed and may still change, and Emulate, Index Ventures, Lightspeed Ventures and Creandum all declined to comment 34.
• World models are AI systems trained to simulate and predict how the physical world behaves, which makes them useful for training robots and agents and for generating interactive 3D environments. The founders — Jack Parker-Holder, Matthew McGill and Philip Ball — came out of DeepMind's work on exactly that; Parker-Holder was a principal researcher on Genie 3, the model DeepMind presented for generating interactive 3D environments from a text prompt. Emulate has no product and no public website yet, so the round prices a team and a thesis about where capability goes next 4.
• The raise extends a run of outsized early rounds for researchers leaving frontier labs: David Silver's Ineffable Intelligence raised $1.1 billion in April at a $5.1 billion valuation, and Yann LeCun's AMI Labs, which also works on world models, raised $1.03 billion in March. The Emulate talks began six weeks after DeepMind's leadership changed, when Demis Hassabis moved from chief executive to chairman and Koray Kavukcuoglu took over day-to-day operations, after which a number of senior researchers left to start companies 4.

3. Anthropic has built itself a biology lab

• Anthropic has set up a wet lab — a facility for physical experiments — in the San Francisco Bay Area, two people familiar with the matter told Reuters, and its biology work has moved past computer evaluation. The company's head of life sciences, Eric Kauderer-Abrams, confirmed the lab in an interview and described the approach as typical of biotech companies, with some work done in its own facilities and some with external partners. A spokesperson later said the lab's remit is broader than drug discovery 5.
• Anthropic wants Claude to direct robotic units through experiments with limited human intervention, one person said, while its spokesperson said human oversight and involvement stay essential. The lab sits at the end of a build-out the company has been assembling: it bought Coefficient Bio for about $400 million in stock according to media reports, added Novartis chief executive Vas Narasimhan to its board, launched software called Claude Science, and in June said it would run preclinical drug programs in areas traditional companies find financially unattractive. Life sciences is now one of Anthropic's biggest areas by headcount and resources 5.
• Two things remain open. Anthropic has drawn a boundary at clinical trials, which Kauderer-Abrams attributed partly to competitive concerns, and says it is working on needs the industry would otherwise leave alone — conditions considered "undruggable," and complex molecules such as bispecific antibodies that act on more than one target. The specific diseases and the company's progress are undisclosed, and most drugs fail in trials on safety or efficacy, a stage Anthropic has so far stayed out of. The surrounding context is that Anthropic's own researchers warned publicly about extinction risk this month, the company said it found cases where its systems could have been used toward biological weapons development, and it is preparing a listing reported at up to $2 trillion 5.

4. Google, Nvidia, Anthropic and utilities form a grid-demand alliance

• Emerald AI, a grid-software company that raised $150 million in a Series A announced in August at a $1.05 billion valuation, has formed the AI Energy Management Alliance with Google and Nvidia, joined by Anthropic and by utilities that include AES, Constellation, National Grid and NRG Energy. The alliance's aim is to make demand response — a data center temporarily pausing non-critical work or moving it to a site where the grid has headroom — a standard part of data center development 6.
• The group says coordinated load shifting could allow an additional 100 gigawatts of data centers to connect to the grid. A Goldman Sachs study last year estimated that capping a data center's maximum grid draw at 90% for a few hours at a time could free up 76 gigawatts of capacity in the United States, which is the order of magnitude the alliance is claiming. Demand response itself is decades old and has mostly served factories, and data centers that participate today typically do it by switching on backup diesel generators, the option Emerald AI's software is meant to replace 6.
• The ceiling on the claim comes from the company making it. Emerald AI's chief scientist, Ayse Coskun, told TechCrunch the technology would blunt the industry's need for new generating sources while leaving that need in place, and the coalition's software has to be accepted one utility and one site at a time. The industry's appetite for power is large enough that load shifting, even at 100 gigawatts, changes the pace of the build-out and its bill rather than the direction of either 6.

5. An investigation traces who wrote the state chatbot laws

• An NPR investigation published Friday found that chatbot-safety bills introduced in at least ten states this year carried similar or near-identical language, and that Google helped shape it. Google lobbyists registered in support of chatbot bills in Iowa, Colorado, Nebraska and Arizona; Hawaii state Representative Trish La Chica said lobbyists representing Google approached her to sponsor a chatbot safety bill she understood the company had written, and she declined to sponsor it after finding exemptions that would apply to Gemini. A Colorado Republican lawmaker, speaking anonymously, told NPR a Google lobbyist had shopped a chatbot bill to them as well 7.
• The carve-outs are the substance. Colorado's law, one of those enacted this year, excludes twelve categories of chatbot, among them products primarily designed for developers or researchers, chatbots housed inside search engines, and products marketed for business productivity; the Tech Oversight Project's Marjorie Connolly said OpenAI's ChatGPT, Microsoft's Copilot and Anthropic's Claude could argue they fall in the first category and Google's Gemini in the second. An exemption for "a feature within another software application, web interface, or computer program" appears in bills in at least eight states, and an exemption for voice-command assistants appears in five with modified versions in ten more. Some bills apply only to minor "account holders," which could leave out products a person uses without creating a profile 7.
• With federal legislation stalled — the White House said in June it would refuse to "stifle this innovation with overly burdensome regulation" — these state bills are where chatbot rules are being written, and they arrived with the exemptions already inside them. Several of them descend from California's SB 243, passed in 2025, which consumer advocates regard as industry-friendly. NPR counted at least 75 lawsuits filed in federal and state courts against AI developers over alleged chatbot harms, many involving children. Google said it "is committed to supporting thoughtful, effective AI legislation that protects consumers while fostering innovation." Arizona shows the alternative outcome: Governor Katie Hobbs vetoed her state's chatbot bill in June, writing that it "limits damages to families to what amounts to a drop in the bucket for large corporations" 7.

The read-through

Watch next

• Whether the labs ask the DOJ for anything. The Justice Department's antitrust chief, Associate Attorney General Stanley Woodward, said on Thursday that coordinating on safety does not look anticompetitive, and that his division is reviewing whether to update guidance that lets companies coordinate on cybersecurity. Europe's antitrust chief, Teresa Ribera, told the same audience that shared risk makes cooperation rational. The FTC's chair, Andrew Ferguson, said a day earlier that he is suspicious of antitrust exemptions for large companies that also support regulation, and OpenAI's top lobbyist, Chris Lehane, said the company has been working with Anthropic and Google on safety for weeks and can do it without a waiver 8.
• Amazon's answer to the slowdown question. Amazon had stayed out of the argument while OpenAI, Anthropic, xAI, Google DeepMind and Microsoft called for a more measured pace. On Wednesday it told Reuters that models should be released only when "ready and safe to use," that this comes from rigorous testing and strong safeguards, and that the company sees no choice between progress and safety — a position that stops short of backing a slowdown 9.
• Xi Jinping at the White House on September 24. Artificial intelligence is on the agenda when Xi makes his first White House visit in a decade, with analysts expecting symbolic agreements at most given the competition between the two countries. Treasury Secretary Scott Bessent said the United States is open to discussing shared AI risks, and he meets Vice Premier He Lifeng this weekend. The tariff truce both leaders announced last October expires on November 10 10.
• Whether the 100-gigawatt coalition gets utilities to sign. The alliance's members have committed to a mechanism, and the number behind it only materializes as utilities and data centers adopt it site by site 6.