Seven X signals: Anthropic measures its own R&D, OpenAI's legal GPT-6, and a benchmark audit

In the 24 hours ending September 18, 2026, the day's biggest announcements came from labs describing their own systems. Anthropic published a way to measure how much of its research Claude now does, and opened a vetted route into biology for its models; OpenAI wrapped GPT-6 Astra in a legal search index and put it in front of law firms; and the sharpest check on the day came from a group auditing everybody else's benchmarks. Ten original posts fell inside the window; they make the seven entries below, grouped into four clusters. Retweets, small talk, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked.

Anthropic proposes three ways to measure the pace of AI development, and publishes its own numbers On September 17 Anthropic published a method for measuring three things inside a frontier lab: how much of its AI research and development the AI itself performs, how well the agents working on its systems are overseen, and where its compute goes 12. As of August, Claude "leads" 26% of Anthropic's measured AI R&D work, up from under 1% in February, and more than 90% of that work sits at or above "AI collaborates". About 30,000 agents run on its most-used internal platform at any moment, where 0.002% of more than a billion decisions were blocked in real time and roughly 50 flagged transcripts a week reach human review. Over one week in July, 6% of the compute going to AI R&D, and 12% of the compute used by research agents, went to safety work 1. Anthropic measured itself, with its own models as judges: the judge matched human raters exactly 59% of the time against 35% agreement between the humans, the compute split covers a single week and rests on best-effort workload labels, and the company calls its definition of safety work one reasonable choice among many 1.

Anthropic opens a verified route into biology, with Mythos included for the first time Under the Life Sciences Verification Program, teams that clear a review of their research credentials, security standards and ethics oversight get Mythos, Opus and Sonnet with safeguards more permissive for biology than the generally available models, which covers drug discovery, research biology, clinical development and manufacturing; dozens of organizations came through an early-access round first 34. Standard Use grants cover a whole team's daily work and renew each year. High-risk Use grants remove the safeguards that block life-science requests for a single project and renew every six months, and for Mythos they stay limited to a small set of vetted entities while Anthropic works with the US government on wider availability 3. Access is tied to the use cases an applicant declares, monitoring shifts from blocking requests in real time to reading traffic afterwards, and program traffic is retained for 30 days to make that review possible. The risks the program is built around — malware diverting an account, an insider acting badly, agents running long-horizon swarms — are ones Anthropic describes, citing its own threat report 3.

GPT-6 Astra gets a legal search index and a route into law firms Astra for Law is a configuration of GPT-6 Astra: the same model carrying legal instructions, settings and a search index of more than 230 million URLs covering US case law, statutes, regulations, court rules and administrative decisions, with the case law drawn from the nonprofit Free Law Project's CourtListener database and sold with 26 partner-built plugins and 47 community skills for work in ChatGPT 5678. The benchmark OpenAI chose on awards finding the right authority and the right passage inside it, and on OpenAI's self-reported runs Astra for Law passed the all-pass check on 54% of 200 questions from the private validation set of Vals AI's Legal Research Bench, against 38.7% for GPT-6 Astra with web search at its highest reasoning setting; Vals AI's own accounting of those same runs gives 53.7% against 39.6% 5910. Those figures come from runs the vendor performed and reported rather than from the benchmark's own harness, where the leading models sit at 55.29% all-pass 910. Access starts through a Trusted Access programme in ChatGPT and Codex for selected firms, an API version is promised without a date, and Mollick's question is whether the labs simply absorb every valuable AI vertical now that they set token prices and own the models underneath 511.

Epoch AI starts grading the benchmarks everyone else quotes Epoch AI, a research group that tracks AI progress, opened Benchmark Reviews on September 17 to audit external benchmarks independently. The first batch covers 15, of which 4 come back Verified, 9 Flawed, and 2 with too little information to review 1213. The labels come with definitions a reader can apply to any number they see quoted: Verified means the benchmark can broadly be interpreted as described and its errors do not substantially affect results; Flawed means it carries one or more substantive flaws that users need to know about to read its results accurately; and Epoch will not review benchmarks it built itself 13. Mollick's summary is that Epoch does some of the best public benchmarking work on AI, and that the exercise also shows how poor the state of benchmarking is 14. The rubric is Epoch's own, the sample is 15 benchmarks at one version each, and a single review such as Terminal-Bench 4.0.0 sets out the rubric and the evidence behind its verdict 1315.

The hard part of running many agents turns out to be smaller than expected Mollick says he overestimated how hard it would be to orchestrate large numbers of agents, having assumed that building working organizations of them would need research of its own; in his experience they self-organize well, and politely 16. His example is a Fable coordinator inside Claude Projects passing messages between agents. If that holds generally, the scarce work moves from wiring agents together to deciding what the organization is for and checking what it hands back. The post carries no task count, cost or failure rate, and the screenshot shows the coordinator rather than the result, so it is a practitioner's revised estimate 16.

A 33,000-volume library rebuilt in 3D, then released as open source Mollick used the new Claude Projects to reconstruct Umberto Eco's 33,000-volume library from images, video and records about it, book by book and shelf location by shelf location, and published the result as an interactive 3D map 1718. He then put the underlying library on GitHub under an MIT licence, with the note that the AI found all of the references by itself 1920. Neither the post nor the repository verifies the reconstruction: the repository carries no description and no tests, and had three stars when it was read, so how closely the shelf locations match the records is unestablished 20.

Nonprofit leaders describe staff resistance that is argued on energy grounds Mollick says nonprofit leaders he talks to describe widespread resistance from their staff, usually argued on environmental grounds that mix real objections with invented ones 21. Environment and energy are the frame his sources use, and the frustration he hears is from leaders whose organizations are stretched thin for their mission and who see AI as a way to do more with what they already have. His evidence is those conversations: no survey, no sample and no rate, so "widespread" is his word for what he is told 21.

Three of the seven entries are organizations describing or selling their own systems — two Anthropic publications and OpenAI's legal offering — and the two sharpest numbers in the issue, 26% of AI R&D work led by Claude and 54% on a legal research benchmark, were produced by the organizations making the claims. The one audit here is of other people's benchmarks and other people's agents. Every entry links the page that carries its number, so the check is available to anyone who needs it before repeating the figure.