Anthropic Says Opus 5.5 Cuts Costs by 40%. Here's What That Number Actually Measures.

Anthropic says Claude Opus 5.5 produces output more than 30% faster than Claude Opus 5 and cuts "typical workload" costs by 40% — released September 22, 2026, alongside higher subscription usage limits and a reported 57.8% score on CursorBench 4.0 at maximum effort. If these vendor-reported numbers hold up in real production use, they'd meaningfully affect the cost and responsiveness of coding agents that make repeated model calls.

Here's the catch: the speed claim is about output generation, not full task completion. In an actual agentic coding workflow, real-world wall-clock time depends on far more than raw output speed — repository setup, tool calls, test execution, retries, context handling, and the number of steps needed to land a correct change all factor in. Anthropic hasn't published how it calculated either the speed or cost figures.

The pricing details are concrete: Opus 5.5 offers a one-million-token context window at $4 per million input tokens and $20 per million output tokens, according to a model-overview page. But those raw prices don't automatically add up to a 40% workload cost reduction — total spend depends on token volume, output length, retry counts, context reuse, and how often the model actually succeeds on the first attempt. A model could plausibly reduce spend by needing fewer retries to complete work correctly, but there's no production data here demonstrating that's actually happening.

The CursorBench score deserves the same caution. It's a company-reported benchmark figure, not independent evidence of performance across private repositories or real long-running maintenance work. There's no published methodology, model configuration detail, comparison baseline, independent replication, or data on how the model handles tool failures and recovers from errors.

Bottom line: Anthropic has made a specific, testable set of efficiency claims — which is more useful than vague marketing language. But whether the 30% speed gain and 40% cost reduction actually generalize across different codebases and engineering teams is still unproven. Teams evaluating Opus 5.5 will need to measure task success rate, completion time, retries, and total cost in their own workflows before trusting the headline numbers.