Claude Opus 5.5 price cut fails to lower cost per task, Artificial Analysis finds

The two announcements came about 100 minutes apart on September 22, 2026: Anthropic announced Claude Opus 5.5, and OpenAI followed with GPT-6 Sol and GPT-6 Luna. The article frames this as a story about where in the cost equation each company moved, not simply about cheaper AI. Anthropic also reflected the cut in the usage limits of its Pro, Max, and Team plans, saying users can work about 25% longer than with Opus 5. OpenAI cut GPT-6 Sol to $2 input and $10 output per million tokens, half of GPT-5.6 Sol's $4 and $20, though that prior price was itself a three-month promotional price introduced on August 21, according to press reports.

Artificial Analysis data shows the two cuts landed differently. GPT-6 Sol's cost per task fell to $1.06 from $1.99, while its output tokens per task rose slightly, from 29,000 to 31,000. Opus 5.5's cost per task in its max setting was $5.98 versus $5.86 for Opus 5. The reason lies in the volume of tokens used. Anthropic has said Opus 5.5's standard setting was moved one step lower, and its 40% saving applies at that standard setting, not at max. Anthropic also published its own example in which moving a 44-query support evaluation to Opus 5.5 at the low reasoning setting cut costs by about 18%, and trimming prompts by another 9%, for about 25% below the Opus 4.8 baseline. The company itself calls this one evaluation result, an example rather than an expectation.

The deeper question is what happened to the provider's own cost, which never appears on a customer's bill. Anthropic said Opus 5.5 needs less compute to serve than Opus 5, but that sentence appears only on its announcement page and not in its 230-page system card, which also does not disclose the model's parameter count. OpenAI said caching and inference improvements let it serve models at lower cost, without giving numbers, though in July it had described efficiency work across model, inference, and agent layers. DeepSeek and Moonshot AI disclosed their architectures in model documentation. The outcome for buyers appears to hinge on whether providers keep disclosing enough to let them compare, and on whether users measure cost per task rather than per token.