Anthropic's AI speedup: Inside METR's findings
METR says Anthropic's progress roughly doubled in 11 weeks, mostly thanks to its own AI.
What METR found inside Anthropic
What this means for developers using frontier models
The bottleneck question and the road ahead
## What METR found inside Anthropic
METR spent 11 weeks with Anthropic's internal data and staff interviews. Its estimate: Anthropic's rate of AI progress is running about 2x faster than its 2024–25 pace, within a 1.3x to 3x range, with AI contributing roughly 80% of the acceleration. The report also surfaced internal models beyond the public Claude Opus 5.5, including a "Model 4" that hit 50% on all 15 tested benchmarks. Anthropic redacted the strongest evidence, citing the need not to hand competitors a roadmap.
## What this means for developers using frontier models
The immediate takeaway for engineers is planning uncertainty. Faster progress means your application can suddenly get better or break in unexpected ways when Anthropic ships its next model. Keep a model-call abstraction layer in place, maintain a solid evaluation suite, and test new versions before they reach production. The benefits cut both ways: AI-assisted coding tools will also absorb some of this internal speedup, so workflows that were tedious six months ago may be dramatically cheaper soon.
## The bottleneck question and the road ahead
METR reports no clear bottlenecks, suggesting the self-improvement loop could continue compounding. Yet Anthropic's own risk language treats "dramatically accelerated progress attributable to AI" as a danger threshold, and METR says that point may be close. For developers, the smart operating assumption is that today's frontier won't remain stable for long. Invest in adaptable architectures and continuous benchmarking rather than trying to lock onto one model version.
If you build on frontier AI, plan for a faster model landscape than yesterday's roadmap assumed.