Anthropic invites philosophers to sharpen Claude's reasoning

Anthropic is now recruiting philosophers, economists and lawyers as well as engineers. Each fellow gets four months to help Claude reason well about the hardest problems in their own field. According to Anthropic's announcement on X, applications for two new fellowships are open until October 28, and the work starts in January 2027.

At a glance

• Anthropic has split the new fellowships into two tracks: STEM Fellows for science and engineering, and Conceptual Reasoning Fellows for philosophy, cognitive science, economics, policy and law.
• Both groups spend four months at Anthropic from January 2027. The existing Fellows program pays $3,850 a week plus roughly $15,000 a month in compute, a likely but unconfirmed benchmark.
• The announcement gives no cohort size, pay or location for the new tracks. The closest reference point is an older fellowship built around empirical safety research.

If you have not followed it, Anthropic already runs a Fellows program for AI safety research. According to Anthropic, more than 200 fellows have taken part since early 2025, and over 65% have published work from their fellowship, including Petri, subliminal learning and open-source circuits. Open-source-ai reports that the 2026 program added tracks for ML systems and performance, reinforcement learning, and economics and societal impacts.

The two new tracks close applications on October 28 and start in January 2027

The STEM track covers science and engineering. The Conceptual Reasoning track covers philosophy, cognitive science, economics, policy and law. Fellows on both tracks spend four months at Anthropic from January 2027, and their brief is the same: to help Claude reason well about some of the hardest problems in those fields.

Each track has its own listing on Anthropic's Greenhouse job board, one for Conceptual Reasoning Fellows and one for STEM Fellows. Both close on October 28. The January cohort of the existing Fellows program closes on October 18, so the two programs have different deadlines and you should not mix them up.

The existing Fellows program pays $3,850 a week plus roughly $15,000 a month in compute

The announcement says nothing about money, so the nearest guide is the existing program. Its fellows get a weekly stipend of $3,850 (£2,310 or $4,300 CAD) plus roughly $15,000 per month in compute. The compute is budgeted separately because most empirical safety work needs real model access at scale. Nobody has confirmed either figure for the STEM or Conceptual Reasoning tracks.

Earlier cohorts show where fellows end up. Over 80% produced papers and more than 40% went on to join Anthropic full-time. Another listing puts full-time offers in earlier cohorts at 25 to 50%. The pilot cohort had 10 to 15 fellows, and later cohorts have not published a fixed size. According to Open-source-ai, the program competes for candidates with MATS, the Astra Fellowship and academic post-docs.

Fellows will spend four months working on Claude's reasoning, starting January 2027

The announcement does not describe the day-to-day work. Anthropic does describe it for the existing program. Mentors pitch projects, and fellows choose and shape them in close collaboration with those mentors over four months, aiming for public outputs such as a paper. It works a bit like a residency: senior staff bring the list of problems, and the fellow picks one to own.

The practical terms come from the listing for the existing January 2027 cohort, as described by DevGrants Daily. It is full-time with a possible extension. Fellows work from Berkeley or London, or remotely from the US, UK or Canada. Applicants need work authorization in those countries because Anthropic does not sponsor visas, and a PhD, prior ML experience or published papers are not required.

According to the same listing, selection goes through an application and reference check, then technical assessments, interviews and a research discussion. Neither the announcement nor the listing says whether the philosophy or law track keeps the technical assessment.

The announcement does not say what a philosophy or law fellow should have produced by the end of four months. In our view, the odder gap is selection. DevGrants Daily lists technical assessments for the existing cohort, and nothing explains how that step applies to an economist. Open-source-ai also raises a lock-in concern: fellows work on Anthropic models with Anthropic mentors, which biases technique selection toward problems tractable on that stack.

Before the October 28 cutoff

Applications for both tracks close on October 28, and fellows start in January 2027. Anthropic has not said how many people each track will take, what they will be paid or where they will work. The existing program reviews applications on a rolling basis and runs several cohorts a year. Nothing says yet whether the STEM and Conceptual Reasoning tracks will run again after January.