Anthropic wants to fund a new generation of AI benchmarks

Today's AI benchmarks are fundamentally broken.

Most tests cited across the industry were built before modern generative AI even existed. They measure narrow, academic tasks rather than how real people use models, and they completely fail to capture catastrophic risks.

Now, Anthropic wants to fix that by footing the bill. The Claude maker just launched a funding program to sponsor third-party organizations that can build a new, comprehensive generation of AI benchmarks.

Beyond multiple-choice tests

Anthropic says high-quality safety evaluations are in severe shortage across the ecosystem. Applications will be reviewed on a rolling basis, with customized funding options for different project stages.

Rather than standard coding or reasoning riddles, Anthropic wants to evaluate frontier threats. That includes testing a model's capacity to execute cyberattacks, assist with weapons of mass destruction like nuclear devices, and deceive humans via deepfakes or misinformation.

They are also pledging to build an early warning system for national defense threats, alongside tests for end-to-end scientific research, multilingual abilities, and bias mitigation.

To make this work, Anthropic is planning platforms where domain experts can build custom evals and run trials with thousands of real users. Selected teams won't just get cash; they will work directly with Anthropic's frontier red team, fine-tuning, and trust and safety units. Anthropic even hired a full-time coordinator and noted it may outright acquire or expand promising projects.

Who controls the scorecard?

While funding open research sounds noble, Anthropic's dual role as a commercial frontier lab is raising eyebrows.

In its announcement, Anthropic acknowledged that funded evaluations must align with its own safety classifications, developed alongside groups like the nonprofit METR. That essentially requires grant recipients to adopt Anthropic's proprietary worldview of what counts as safe or risky.

There is also growing pushback from researchers tired of existential risk talk. Critics argue that hyping up Hollywood-style nuclear threats distracts regulators and the public from everyday AI harms already happening right now, like persistent hallucinations and systemic bias.

Why it matters

In our view, Anthropic is pulling off a brilliant regulatory and platform play.

Traditional LLM benchmarks have hit a dead end, with every major model scoring in the high nineties. By bankrolling the next generation of evaluations, Anthropic is effectively shaping the ruler that will measure the entire frontier AI race, putting pressure on OpenAI and Google to submit to external safety rubrics.

For independent researchers, the funding and direct access to top-tier red teams are genuine lifelines in an underfunded space. If you are building rigorous eval infrastructure, applying is an absolute no-brainer.

However, the catch is obvious. Independent benchmarks only carry weight when they are completely untangled from corporate interests. Taking money from a commercial lab that answers to investors means walking a very fine line between advancing safety science and reinforcing a single vendor's standard.