METR: Anthropic's AI development pace has roughly doubled, driven by AI-assisted research

METR, the independent group that evaluates frontier labs for dangerous capabilities, concluded that Anthropic's development pace has roughly doubled year over year. Its finding is that most of the acceleration comes from AI systems helping with Anthropic's own research and engineering. This is an early, empirical version of the recursive-improvement loop that safety researchers have long warned about.

The method relies on internal productivity data, research throughput and interviews, not public benchmarks alone. Much of the key evidence is redacted, and critics on dev.to and elsewhere pointed this out quickly. Anthropic co-founder Sam McCandlish was cautious, saying AI is speeding up research but that "it's hard to say exactly how much."

The financial context matters too. Anthropic's Q3 revenue grew 47% to $17 billion. The company still posted a $567 million operating loss, which it attributes to costly short-term compute rentals. A record IPO is expected later this month, so the METR report doubles as a capability signal and an investor narrative. It also feeds the wider debate about AI labs speeding up while politicians, including an Arizona congresswoman whose call for a pause went viral on Reddit, push for slowdowns.

What to watch: whether METR or Anthropic releases less-redacted evidence, whether other labs face the same evaluation, and how IPO disclosures describe compute costs. A doubling pace paired with an operating loss is the clearest sign yet that frontier progress and its margins are moving in opposite directions.