Anthropic released Claude8.3 out of 10 Opus 5 on July 24, 2026, positioning it as a step change for its top-tier Opus line aimed at long-running agents, coding, and professional work, while holding the price of the model steady rather than raising it.
Pricing and specs unchanged, capability claims up
Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — identical to the outgoing Opus 4.8 — with a faster mode available at roughly 2.5x the response speed for twice the base price. It ships with a 1-million-token context window and is now the default model on Claude Max. A new “effort” dial lets developers choose between low, medium, and high reasoning effort per request, giving enterprise customers a lever to manage inference cost against task difficulty.
Anthropic says Opus 5 sets new highs on its internal and third-party coding and knowledge-work evaluations. The company points to more than double Opus 4.8’s score on its Frontier-Bench v0.1 evaluation, a result on CursorBench 3.2 at maximum effort that lands within roughly half a percentage point of Anthropic’s higher-end Fable 5 model at half the per-task cost, and a score on the ARC-AGI 3 novel-problem-solving benchmark roughly three times the next-best model cited.
Why it matters
The launch continues 2026’s pattern of frontier labs competing as much on price-per-capability as on raw benchmark leadership: OpenAI shipped its tiered GPT-5.6 family (Sol, Terra, Luna) earlier this month, and Google pushed out three new Gemini 3.x variants in the same window. By holding Opus pricing flat while claiming a large jump in coding and agentic benchmark scores, Anthropic is betting that cost-conscious enterprise buyers — many of whom already route agentic workloads like Claude Code8.3 out of 10 through Opus-tier models — will treat Opus 5 as a straightforward upgrade rather than a new spending decision. The effort toggle in particular targets a common enterprise complaint about agentic AI: unpredictable per-task inference costs at scale.