Artificial Analysis Launches New API Performance Benchmark Centered on Time per Task

On November 5, 2026, Artificial Analysis announced the next version of its API Performance Benchmarking on X, making Time per Task the headline performance metric. Instead of how quickly a model streams tokens, the benchmark now measures how long a model takes to finish an entire task. The update covers roughly 1,000 serverless endpoints in Artificial Analysis's testing.

In the same announcement, Artificial Analysis noted that most leading labs set their API default below the highest inference-effort tier, which sharply reduces Time per Task. GPT-6 Astra, for example, scores 52 on the Intelligence Index at medium effort versus 57 at max effort, but finishes a task in about 5.8 minutes instead of roughly 18 minutes. GPT-6.1 Sol drops from 55 to 46, though the piece does not specify which effort level that comparison refers to.

The update also changes how endpoints are scored on speed. Failed requests now count against an endpoint: errors, rate limits, and timeouts receive a speed score of 0 and are included in the endpoint's rating. Each endpoint now displays an Endpoint Error Rate, so endpoints that fail frequently drop directly in the speed ranking. That is where the published information stands at this point, with no further developments reported beyond these metric changes and the accompanying model data.