Anthropic Fellows May 2026 cohort

The Anthropic Fellows program runs two 2026 cohorts beginning in May and July, each a four-month empirical research fellowship paying a weekly stipend of $3,850 USD (£2,310 GBP / $4,300 CAD) plus roughly $15,000 per month per fellow in compute. Fellows work under Anthropic researcher mentorship with the explicit deliverable of producing a public output, typically a paper.

The 2026 cohort tracks span scalable oversight, adversarial robustness and AI control, model organisms of misalignment, mechanistic interpretability, AI security, and model welfare. Anthropic also added parallel fellow tracks for ML systems and performance, reinforcement learning, and economics and societal impacts, branching the program from a pure alignment fellowship into a broader research pipeline. The pilot cohort was 10 to 15 fellows; later cohorts have not published a fixed size, but over 80% produced papers and more than 40% subsequently joined Anthropic full-time, with others going to external safety organizations.

Past cohorts have shipped work on agentic misalignment, subliminal learning, rapid response to ASL-3 jailbreaks, and open-source interpretability circuits. The compute allocation is sized as a separate line item from the stipend because most empirical safety work needs real model access at training-run scale, which a typical post-doc budget does not cover.

In the funding context for alignment researchers, the program competes with MATS, Astra Fellowship, and external academic post-docs. Its distinguishing feature is bundling researcher pay with material compute in a single grant from the lab whose models are being studied, rather than a host institution paying salary while the researcher negotiates compute access separately. The lock-in concern is the obvious one: fellows work on Anthropic models with Anthropic mentors, which biases technique selection toward problems tractable on that stack.