Anthropic released Opus 5 with benchmark gains that the company presents as incremental rather than a new capability tier. The model is priced like its predecessor and is positioned below the company's most capable cyber-focused systems.
Anthropic released Opus 5 on July 24. The model is available more broadly than the company's higher-end Mythos offering. Anthropic priced Opus 5 at five dollars per million input tokens and twenty-five dollars per million output tokens.
Company benchmark charts showed modest gains over Opus 4.8 and competing coding models. Anthropic said Opus 5 was not trained to match Mythos 5 on offensive cyber exploitation. The checked analysis described the product as a cost and availability update more than a capability leap.
Model benchmarks depend on task design, scoring and tool setup. Routing work among models can reduce cost when the hardest model is unnecessary. List price does not capture retries, latency, integration or supervision.
The checked record also defines what is not yet established. The benchmark figures came from the vendor, independent production comparisons were not yet available, and actual savings will vary by workload. This distinction prevents an announcement, allegation, estimate or early field report from being presented as a completed or independently proven event.
At the August 2 publication cutoff, the next evidence expected to update this account is independent coding and agent evaluations and measured cost, latency and error rates in production use. Those future developments are not assumed here; they will require a responsible source, a dated public record or independently verifiable reporting.
The source pages retained below support the numerical values, sequence and attributed statements in this report. Statements from interested parties establish what those parties said or did, but they do not independently prove every claim embedded in those statements. No image is included because a rights-cleared visual was not necessary to report the facts.
This edition preserves the difference between the immediate event and its operating context. Model benchmarks depend on task design, scoring and tool setup. Routing work among models can reduce cost when the hardest model is unnecessary. List price does not capture retries, latency, integration or supervision. The article will remain fixed at this cutoff even if a later investigation, corrected total, weather observation or implementation record changes the public understanding.