Anthropic AI Speed Doubles, But METR Evidence Redacted
Anthropic's models are improving ~2x faster, but redacted evidence makes the claim hard to verify.
What METR found inside Anthropic
The missing evidence is the real story
Why the contested pace matters to developers
The takeaway for engineers and policy
## What METR found inside Anthropic
METR's 11-week independent review—conducted without payment and with access to internal benchmarks, staff interviews, and surveys—found that Anthropic's AI progress speed, measured by its internal capability index, is now about 1.9 times the pre-October 2025 rate, with a 95% confidence interval spanning 1.5x to 2.8x. In practical terms, that is roughly two years of historical progress compressed into one. METR attributes the lift primarily to AI-assisted engineering—researchers using models to write code, run experiments, and try more ideas—and says it can confidently rule out faster training-compute growth as the main driver. Staff surveys suggest an informal self-reported speedup of about 12x on the latest internal model, which METR treats as suggestive, not conclusive. METR also heard that AI is not yet making major fully autonomous contributions to the hardest training steps, though signs of life are emerging.
## The missing evidence is the real story
The most important detail is outside the report. Anthropic held unilateral redaction rights and used them to remove most of the evidence METR considered most informative for risk conclusions, including stage-by-stage acceleration estimates, bottleneck analysis, and future-model forecasts. METR is explicit that the public report 'does not achieve transparency about evidence.' The result is not an open audit but a trusted intermediary's opinion: useful, and far better than nothing, but not independently verifiable. For a finding that could reshape risk timelines, that gap matters.
## Why the contested pace matters to developers
If frontier labs are genuinely progressing twice as fast as a year ago, the implications extend beyond research labs. Teams building on frontier models may face faster capability shifts, shorter lead times for safety work, and less time to adapt. The difference between 1.3x and 3x is not academic; it changes how quickly skills and roadmaps may need to move. Yet the number is contested: acceleration could be concentrated in clean, auto-gradable tasks that overstate messy real-world impact, and Anthropic's earlier baseline was slower than the broader industry frontier, which could inflate the 2x comparison. At the same time, redacted internal metrics hinted acceleration could be above 2x and still rising, with no clear bottlenecks in sight. That combination is exactly the recursive self-improvement scenario safety researchers worry about.
## The takeaway for engineers and policy
Treat '2x faster' as a high-uncertainty warning sign, not a proven metric. METR estimates a majority—roughly 80%—of the acceleration is due to AI assistance, but the supporting measurements are hidden. Anthropic says it plans its own public analysis of internal progress speed, which could fill some gaps, but independent scrutiny will remain limited until underlying data is open. External watchdogs and developers both need public, inspectable evidence before adjusting deployment plans, risk models, or regulatory expectations. Until that evidence exists, the honest position is that progress appears to be accelerating, and the public cannot yet judge how much.
Read the 2x headline as a directional signal, not an independently verified fact.