Artificial Analysis Makes Time per Task Headline

This is a necessary correction: streaming speed measures how quickly a model types, not how quickly it completes useful work.

• –Opus reaches roughly 112 tokens per second versus Astra’s 47, but takes 25.8 minutes per task versus 18.4 minutes.
• –Astra’s advantage comes from generating dramatically fewer tokens, highlighting the cost of verbosity and extended reasoning.
• –The new metric better matches coding agents and knowledge-work systems, where latency, reasoning length, and multi-turn execution compound.
• –It remains a derived benchmark estimate rather than a stopwatch measurement of every production workflow.
• –Developers should compare models at their intended API effort level, since default and maximum effort can produce radically different speed, cost, and quality tradeoffs.