Anthropic closed the month by shipping Claude Opus 5. By the published numbers, the model lands within 0.5% of the house's previous peak on a coding benchmark, at half the cost per completed task, with per-token pricing unchanged.
The chosen metric says a lot about where competition went. Cost per completed task, rather than price per token, measures how many round trips the model needs to reach a result. An expensive model that gets it right first time can cost less than a cheap one that misses three times.
Holding per-token pricing while improving cost per task is a way of cutting price without announcing a cut, which preserves margin and still improves the comparison against competitors.
The week's context makes the launch more legible. While a competitor delayed its flagship over insufficient performance in coding and long reasoning, this one shipped leading with a coding number.
For anyone choosing a supplier, the useful takeaway is the method rather than the scoreboard: measuring cost per completed task on your own workload usually reorders the candidates relative to the price list.