A new chart with straight lines from our favorite METR.
The previous one turned out to be saturated (did not contain more complex and lengthy tasks for evaluation), it was not updated for a long time.
This was drawn as a result of the METR team's work at Anthropic (but it's not exactly the same as what Dario called), much more detail here:.
"We provided Anthropic with a detailed report of our assessment, approximately 100 pages long. In the public version, Anthropic designed much of the evidence we found most informative. As a result, the public has to rely on our judgment more than we think is optimal."
Important perspective: Starting in October 2025, the performance of the internal model is growing at about twice the rate of the previous period. The auditors also ensured that this was not due to shedding additional capacity.