AI is now helping build better AI, and this time there’s outside proof. This isn’t a hunch or a hype thread. I just watched a breakdown from AI creator Matthew Berman, who walks through what the independent evaluator METR found after months inside Anthropic.
It gave me chills. Here’s the short version.
🚀 The Key Idea
METR sent a small team of about 2.5 full-time people into Anthropic starting in mid-August. They wanted answers to three questions:
- Has Anthropic’s internal capability growth sped up?
- How much of that speedup comes from AI itself?
- Is anything likely to slow it down soon?
The answers were yes, mostly AI, and not really.
Most METR staff on the project now believe Anthropic is improving its models on well-specified tasks at about twice the rate it did between March 2024 and October 2025. They wouldn’t be shocked if the real number fell anywhere from 1.3x to 3x. The big takeaway is that they estimate about 80% of that acceleration comes from AI. It’s not coming from more hires, better researchers, or a flood of new compute.
📈 What Stood Out
- The curve bent at Opus 4.5. Anthropic’s internal capability index grew in a fairly straight line until late 2025. Then Opus 4.5 arrived, coding got dramatically better, and the line started climbing faster. Mythos kept it going. The expert points out that the acceleration itself seems to be speeding up.
- There are no obvious brakes. METR says it found no clear bottlenecks in algorithmic progress, training compute, experiment compute, or data that would kick in over the next 12 months. That line is what made the author nervous, and I get why.
- Researchers feel it too. Anonymous staff said things like “could definitely endorse a greater than 2x speed up” and “no longer bottlenecked on code.” Benchmarks are filling up so fast that Anthropic keeps having to make them harder. Unreleased internal models already fit the steeper trend.
There’s one more nuance worth knowing. AI still can’t make big, fully autonomous contributions to the most important parts of model training. But researchers described “significant signs of life,” which suggests that could change soon.
⚠️ The Fine Print
This wasn’t the full “employee-like access” program that the labs promised earlier. Here’s what was limited:
- Anthropic could approve or reject any data request METR made.
- Parts of the analysis, including some bottleneck details, were redacted.
- A planned check on how compute is split between capabilities and safety was dropped partway through.
The creator wonders whether the upcoming IPO played a role in that last change. Anthropic said it wanted to keep the work focused on risk assessment and would come back to compute verification later. You can decide for yourself what that means.
💡 Why This Matters
For me, the biggest shift is that “models are improving faster” isn’t just a feeling anymore. A respected outside evaluator with real internal access looked at the data and agreed. If you build with AI, plan for tools that get noticeably better every few months, not every few years.
Watch the full video to see the charts, the researcher quotes, and the author’s take on where this goes next.