A new report from third-party safety researchers at METR, released on Thursday and heavily redacted at Anthropic's insistence, examined internal Anthropic data over nearly three months of "unprecedented access," including employee interviews, Slack polls, and data on unreleased internal AI models. METR said Anthropic redacted a majority of the evidence it found most informative. The report found that over the last six months Anthropic finished training three models not yet available to the public, including one from 2022 and models finished this July, sometime in August or September, and sometime in September or October; METR estimates these internal-only models are helping Anthropic improve its AI capabilities faster. Internal Slack poll data showed staff reporting major productivity gains, with an average self-reported increase of roughly 12x "for the most capable internal model" as of October, though METR cautioned the polls are weak evidence because respondents may not represent all employees or have given answers much thought. Many interviewed researchers said Claude significantly boosted their work, with one saying work is "Now only bottlenecked on ideas & compute, no longer bottlenecked on code." METR said it did not see "observed bottlenecks" that would prevent acceleration from increasing further over the next year.