AIBRIEF
Anthropic's new flagship model costs half per token but 26% less per benchmarked task — the per token math doesn't track real workload cost. This is a development update: read with our prior Opus 5 per token price cut piece and the memory critique.
Editor's note, July 24: This is a development update to our earlier coverage of Anthropic's Opus 5 launch. The per-token price-cut reporting and the memory-critique piece ran earlier today.
Anthropic's Opus 5 lists at half the per-token price of its predecessor — but the per-task math is smaller. Artificial Analysis puts Opus 5 at $2.03 per Intelligence Index task versus $2.75 for its previous top model, a 26% cut, not 50%.
OpenAI's GPT-5.6 Sol (max) runs $1.04 per task and Moonshot's Kimi K3 $0.95, putting Opus 5 above both on workload cost despite the per-token win.
Anthropic's own announcement frames Opus 5 as "close to the frontier intelligence of Claude Fable 5 at half the price," and Artificial Analysis ranks Opus 5 first on its Intelligence Index at 61, one point ahead of Fable. The framing holds at the token level; it softens once you count tasks.
One buyer-facing wrinkle the pricing headlines miss: Opus 5 has no data retention requirement, unlike Anthropic's default policy for other Claude models.
The next test is OpenAI and Google's list-price response: do they hold steady, or match the per-token cut?