Research Artifact

Custom GPU Kernels for Seven Open-Source Science Models

The interesting claim is not the speedup. It is that the speedup ships with a correctness condition: Anthropic reports the optimized models produce identical outputs, which is the kind of check a reader can actually hold the work to.

Updated 2026-09-196 minClaude Mythos 5.1performance engineeringbuild note

Key Facts

Page summary for citation

  • Research Artifact: The interesting claim is not the speedup. It is that the speedup ships with a correctness condition: Anthropic reports the optimized models produce identical outputs, which is the kind of check a reader can actually hold the work to.
  • Evidence trail: 2 sources, including Anthropic, Anthropic Docs.
  • Internal context: related pages include Venus Map, Gallery, Verify.
AnthropicUp to 2.5xreported inference speedupAnthropic reports speedups of up to 2.5 times across seven open-source deep learning models, with identical outputs.Anthropic30-60%estimated GPU cost cutOn genome-wide analyses, the announcement estimates the optimized models cut GPU costs by 30 to 60 percent.AnthropicPublicsource code onlyThe announcement says the work was done using publicly available source code alone, in days rather than weeks.

"it sped up seven open-source deep learning models by up to 2.5 times (with identical outputs)"

Anthropic, Claude Fable 5.1 and Claude Mythos 5.1 announcement - source

What was built

Anthropic reports that Claude Mythos 5.1 wrote custom GPU kernels and added caching of intermediate results for seven open-source protein and genomics models. The stated inputs were the publicly available source code for those models; the stated outputs were faster inference on an NVIDIA H100 with unchanged results.

The announcement frames the value in terms of who can afford this work. Biologists running a model thousands of times across an experiment pay the slow path repeatedly, and the performance engineering that would fix it is usually a multi-week specialist project. Anthropic says the optimizations will be open-sourced.

The build pattern

This is the narrowest and most copyable pattern on the site: pick a workload where correctness is already defined by an existing implementation, let the agent rewrite the hot path, and gate the change on output equality plus a measured speedup. The oracle is free, so the agent cannot quietly trade accuracy for time.

Readers evaluating their own candidates should look for the same shape. Deterministic reference output, a measurable cost, a public codebase, and a rollback that is just reverting a kernel. Those constraints are what make an agent-written optimization reviewable instead of a leap of faith.

  • Oracle: the unoptimized model output, compared for equality.
  • Metric: wall-clock inference on named hardware, plus estimated cost per analysis.
  • Scope discipline: no change to model weights, architecture, or scientific claims.

Evidence and limits

As with the Venus map, all figures here come from the Anthropic announcement. The kernels were not open source as of this source pass, so the speedup, the equality claim, and the cost estimates cannot yet be inspected by a third party.

The surface also matters for anyone hoping to copy it directly. Anthropic describes Claude Mythos 5.1 as trusted-access only rather than a model most readers can select. The generally available comparison point is Claude Fable 5.1, which shares the underlying model with more restrictive safeguards.

  • Evidence tier: vendor announcement; optimizations not yet published at the time of this check.
  • Access limit: Mythos 5.1 is trusted access only, not a model most builders can pick.
  • Re-check trigger: publication of the kernels, which would move this from reported to inspectable.

Primary Sources

Source trail

Links are cited for factual claims on this page. Access dates are kept so future updates can re-check drift.

Choosing a Claude model

Anthropic Docs - official docs - accessed 2026-09-19

Open citation

Current generally available lineup and official guidance on picking between Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5.