Anthropic opens a $5 million grant program for independent wellbeing evaluations

Anthropic published the program in its announcements feed on 25 August 2026. The money is pointed outward rather than at internal safety work: recipients, the post says, “will work fully independently, and will publish their work as open-source projects that any developer can make use of.” Alongside the cash, selected applicants get access to Anthropic’s models and technical support from the company.

The stated problem is measurement. Anthropic argues that most model behaviour can be judged from a single answer, while wellbeing cannot. The post gives the example of a user asking about losing weight: advice on balanced diets and workout routines is ordinary, but if that same user has shown a history of disordered eating, the company says the response “could be inappropriate, and potentially actively harmful”. It also notes that someone in distress may not mention self-harm straight away, so the need for a more careful reply can surface only deep into a long conversation.

What the guidance asks for

• A clear statement of what is being measured, including what counts as a pass or a fail and why it matters.
• Clinical and subject-matter experts involved in the design and the validation, not consulted afterwards.
• Coverage of both failure directions, which the post names as overcompliance and overrefusal.
• Scenarios that reflect how people actually use these systems, usually meaning multi-turn conversations where risk grows and context shifts.
• Graders checked against real subject-matter experts before their scores are trusted.

The numbers

The published figure is $5 million for the whole program. There is no per-grant ceiling, no target number of awards, and no split between cash and compute. The application deadline is September 21 and notifications for full proposals go out by October 5, roughly a fortnight apart, which is a short window for assembling a clinical collaborator and a methodology.

The other half of the offer is model access, and our Anthropic price table gives it a value. Claude Sonnet 5 is billed at $2 per million input tokens and $10 per million output. Claude Opus 5 sits at $5 and $25, Claude Fable 5.1 at $10 and $50, and Claude Haiku 4.5 at $1 and $5 per million tokens. Cache reads, which is what you pay to resend context a model has already seen, run at $0.20 for Sonnet, $0.50 for Opus, $0.25 for Fable and $0.10 for Haiku per million tokens. Batch work is half price. Claude 4.6 and later carry the full 1M token context at standard pricing, so a long transcript does not push a benchmark onto a separate rate.

Those lines matter because of what the guidance asks for. A multi-turn evaluation resends its whole history on every later turn, so the cost of one scenario grows with the square of its length rather than in step with it. A benchmark meant to be rerun against each new model release, which is the only kind worth open-sourcing, pays that cost again every time. Cache reads and batch pricing are therefore not a footnote in a budget here; they are the difference between an evaluation a small team can maintain and one it can afford to run once.

The open question the announcement leaves is what happens when an independent benchmark reports poorly on the funder’s own model. Anthropic has committed to independence and to open publication, which is the right structure, but the test of it will be the first unflattering result rather than the grant terms.