Costs

What it costs.

Work	Price to the user	Terms
Launch evaluation of a frontier open model	Free to the lab	$200k all-in per evaluation, funded by philanthropy
Compute for third-party evaluations of frontier open models	$1.55 per H200 GPU-hour	Set below published H200 rates
Academic, safety hub and independent safety research	At or below cost	Allocation depends on the proposed work and available capacity
Incident investigations	By arrangement	Supported with GPUs when capacity allows

Compute is billed per H200 GPU-hour. Access, scope and capacity are agreed before work begins, and pre-release launch evaluations take priority during release windows.

37% below the median on-demand H200 rate of $2.47 per GPU-hour on September 28, 2026 (GPU Finder, 31 single-GPU listings). The cheapest H200 in stock that day was $2.00 at PrimeIntellect; Runpod Secure Cloud listed $4.59.What AI safety compute should cost ↗

Launch evaluations

How we assess a model for $200k.

A substantial campaign includes benchmarks, replication, targeted red-teaming, diagnosis and retesting. We budget $200k per major model launch, with smaller iterations being considerably less expensive.

We target three weeks from agreed kickoff through delivering findings and scoped retesting.

We record GPU hours, queue time and cost per completed evaluation and release these metrics after anonymizing the model and lab we worked with. We are always thinking about where the constraints are greatest (hardware, lab endpoints, researcher time, eval complexity) and work to address these bottlenecks.

See the campaign sequence

One scoped evaluation

Evaluation compute and environmentsIncluded
Research and delivery staffIncluded
Investigation, reproduction and scoped retestingIncluded
Storage and data transferIncluded
Allocated operations and contingencyIncluded
Private findings and resource statementIncluded

All-in evaluation price$200k

Model evaluation cost is influenced by model size, eval complexity and a host of other factors, most of all retesting after interventions to make the model safer.

Risks and Cautions

Compute is fungible. Free evaluation can release a lab’s own budget for capability work. Usage restrictions and monitoring reduce particular risks; they cannot eliminate that indirect effect.

Evaluation can also help a lab improve its models. We accept that trade-off where we expect the safety benefit to outweigh the capability benefit. Our case is that these models are likely to be released anyway, and that additional scrutiny before release can expose risks while there is still time to respond.

Trust, safety and security

Let’s scope a grant.

We can walk through the campaign plan, budget and remaining technical questions.

Discuss funding Read the funder brief