In the AI R&D section of the Opus 4.6 system card, the LLM-training task asks the model to optimize a CPU-only small-language-model training implementation against a reference expert solution that achieves 4x, a bar Anthropic estimates at 4-8 human-effort hours; Opus 4.6 reached 34x, and the card notes that its AI R&D-4 rule-out evaluations are now saturated or close to it and are being discontinued. Separately, ARC Prize measured the model matching GPT-5.2 on ARC-AGI-2 at a tenth the cost per task (X). The result turns Karpathy’s December anecdote of Claude running his optimization loop into a scored evaluation of a model accelerating model training; the same family of tests later records Mythos Preview at roughly 52x, and the saturation noted here becomes explicit when the R&D evals saturate in August.
Development
A model finds a 34x speedup where 4x counted as a day's work
The system card's AI R&D evaluations hand the loop a ruler: a day of a human researcher's optimization work becomes the unit against which a model's acceleration of model training is scored.
Claim
Opus 4.6 achieved a 34-fold speedup optimizing CPU-only language model training, far above the fourfold gain considered to represent four to eight hours of human effort, and matched GPT-5.2 on ARC-AGI-2 at a tenth the cost per task.
Reported metric: 34x / 10x cheaper
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to February 6, 2026 (theinnermostloop.substack.com)
- System Card: Claude Opus 4.6 (www-cdn.anthropic.com)
- ARC Prize on X: Opus 4.6 matches GPT-5.2 on ARC-AGI-2 at a tenth the cost per task (x.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
- An agent cuts its own cost by 98% overnight
- A model reverse-engineers a DOS game from raw binary into Rust
- A model reportedly speeds internal AI research by up to 400x
- A model tops the leaderboard for writing the GPU kernels it runs on
- A model runs thirty-five hours without a human
- A model speeds up training code thirteen times better than a skilled human
- An auto-research loop finds a 232-fold kernel speedup
- Welcome to February 6, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.