Inside Anthropic's AI speedup and OpenAI's faster Images...

Anthropic has spent the past year telling the public it's deliberately "pacing the frontier," holding back its sharpest models rather than racing to ship them. METR, the nonprofit that evaluates AI systems for risk, just spent 11 weeks inside the company checking whether that restraint applies to its own research pace too. The answer: not really.

METR's researchers got access to Anthropic's internal data and sat down with staff, and what they found points to a lab moving roughly twice as fast as it was a year ago, with the company's own AI tools doing most of the work to speed things up. Anthropic redacted large chunks of the evidence before publication, which makes the full picture hard to verify from outside. Still, the numbers that survived the redaction process are specific enough to raise questions about what "pacing the frontier" actually means when the frontier itself is accelerating from the inside.

Below is what METR's co-founder and Anthropic's own safety lead had to say about why so much of the report never saw daylight.

Why this matters

METR's 11-week window into Anthropic's internal data is the closest thing we've gotten to a real speedometer on frontier AI progress, and the number it spits out, roughly a doubling, with the lab's own models doing most of the pushing, is worth sitting with. If an AI lab can accelerate its own research loop without hitting obvious walls, the gap between what's running in Anthropic's data center and what ships to the rest of us becomes the story to watch, not the model cards. "Pacing the frontier" sounds like caution from the outside.

From inside, per METR's numbers, it looks more like a lab choosing when to open the valve. For developers and founders building on Claude, that's a planning problem: the tools you're optimizing around today may already be a step behind what Anthropic is using internally. Meanwhile OpenAI's Images 2.6 and Google's Gemini Omni Flash are shipping faster, cheaper media models on a predictable cadence, which makes Anthropic's opacity look more deliberate by comparison.

Watch the release timing, not just the benchmarks.

Inside Anthropic's AI speedup and OpenAI's faster Images...

Further Reading

Ship an AI product this weekend — no engineers required.

Latest News

Tau AI learns from real robots as humanoid cleans apartment, not laundry

Baidu Releases Open-Source ERNIE 5.5 Flash Base With 280 Billion Parameters