Anthropic Is Using Its Own Claude To Build Better AI 1.9x Faster, Because Why Do Human Work When You Can Just Train Your Replacement

Independent research outfit METR dropped a fascinating look under the hood at Anthropic, revealing that their internal AI development pipeline has nearly doubled its speed since late 2025. Auditors were granted rare access to review internal workflows and chat with roughly 15 staff members, piecing together how algorithms are helping build their successors. Roughly 80 percent of this sudden turbo-boost traces directly back to using AI internally to automate the tedious parts of building smarter models.

The acceleration was measured using a specific benchmark index tracking task efficiency, showing a jump between 1.5 and 2.8 times faster than previous baselines. Engineers reported that Claude handles the heavy lifting of writing and reviewing raw code, setting up complex experiments, parsing messy output data, and stress-testing hypotheses in record time. One researcher estimated that the agonizing lag between an initial brainstorm and an actual working experiment has shriveled by a massive 80 percent, fundamentally shifting human effort away from syntax errors.

Employees noted that the primary bottlenecks are no longer waiting for code to compile or scripts to finish, but rather human imagination and raw computing power. While individual staff subjective estimates ranged wildly—with some claiming a ridiculous 12-fold productivity spike—METR treated those self-reports with heavy skepticism due to the lack of controlled experimental setups. Even with the dramatic code-generation boost, fully autonomous AI model building is still a distant mirage, as current models mostly just accelerate isolated research steps rather than orchestrating the entire pipeline.

METR also failed to spot any immediate algorithmic walls or hardware shortages that could slam the brakes on this acceleration over the next year, though they admitted high uncertainty in that forecast. These findings awkwardly collide with previous public statements from Anthropic co-founder Dario Amodei, who famously urged the industry to hit the brakes on AI scaling until safety protocols actually catch up to capabilities. Adding to the mystery, Anthropic redacted large portions of the data shared with the auditors, citing fierce competitive secrets, which conveniently left independent watchdogs squinting blindly at the final conclusions.

It turns out that teaching machines to code is the ultimate corporate loophole for skipping the coffee break, proving that the fastest way to build the future is just letting the current version of the digital overlord do all the overtime.

Source: METR