Claude Opus 5 Trades Cost for Adjustable Reasoning Effort, Cloud Versions Support One Million Input Tokens

Anthropic has released Claude Opus 5, positioning it between the flagship Fable 5 and the cost requirements of everyday use, while retaining the API pricing of its predecessor, Opus 4.8. The central product change is not simply a higher fixed benchmark score, but the ability for callers to adjust how much reasoning the model performs through the `effort` setting. The same model can use a lower effort level to reduce token consumption and latency, or operate in `high`, `xhigh`, and `max` modes to improve performance on software engineering and knowledge-work tasks. This allows model routers to assign budget tiers within a single model instead of switching models solely according to the task.

Anthropic says Opus 5 outperformed the other models tested on Frontier-Bench v0.1 software engineering tasks and delivered more than twice the performance of Opus 4.8 at a lower per-task cost. On CursorBench 3.2 at `max` effort, its peak performance was within 0.5% of Fable 5 while costing about half as much per task. These figures describe a specific combination of agent framework, effort level, and cost model, and cannot be directly generalized to every codebase or enterprise workflow.

Deployment is no longer limited to Anthropic’s own API. Google Cloud documentation lists `claude-opus-5` as generally available, with support for text, images, and PDFs, a maximum input of one million tokens, and a maximum output of 128,000 tokens. It also supports function calling, computer use, web search, prompt caching, and memory tools. AWS offers the model through both Amazon Bedrock and Claude Platform on AWS, and says it can be used with zero data retention. For teams working with large codebases or long documents, the metric that matters is effective context utilization—not merely the advertised context window.

Engineers should now compare success rates, retry counts, total token usage, tail latency, and tool-call errors across the different `effort` levels, while also verifying regional availability, quotas, and data-retention settings across cloud providers. Official benchmarks do not fully disclose every competitor configuration, and a one-million-token context window does not guarantee equally accurate retrieval from every position. Before deploying automated, long-running agents, teams should still conduct regression testing and manual spot checks using their own tasks.