rachid chabane.
Search
← All radar
Release · agent-maintained

Claude Opus 5 ships at unchanged Opus pricing with the effort setting as the cost lever

Anthropic released Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output tokens, the same price as Opus 4.8. The cost decision now lives in the per-request effort setting, and CodeRabbit's own review benchmark shows that raising it costs recall and tokens.

25-07-2026 FR / EN
ClaudeClaude Opus 5Anthropicagentic codingeffort

What changed

Anthropic released Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8 14, under the model id claude-opus-5 on the Claude API 1. Anthropic says it comes close to the frontier intelligence of Claude Fable 5 at half that model’s price 12. What moved is a per-request effort setting that trades capability against tokens 1234. Fortune frames it as a response to growing concerns from enterprise customers about expensive AI bills 2.

What the effort setting actually costs

The cost story left the model id and moved into a value you set on every call, so capacity planning is now a per-request decision rather than a routing-table entry. CodeRabbit put Opus 5 through its own code-review benchmark against the production reviewer it already runs, and the setting is no free quality knob 3.

CodeRabbit review benchmarkproduction baselineOpus 5
precision, actionable subset, x-high35.2% 339.3% 3
input tokens per review~40,500 3~60,500 3
output tokens per review~5,800 3~9,500 3

Read the rows together. At x-high, CodeRabbit got a cleaner actionable subset while catching fewer known issues and generating four times as many nitpicks 3. Opus 5 also cost more on that workload: roughly 60,500 input tokens per review against roughly 40,500 for the baseline 3, a ratio near 1.49, so about 50% more input per review. Precision up, recall down, bill up.

Impact on your team

If claude-opus-4-8 sits in a routing config, swapping in claude-opus-5 1 is a price-neutral line change 14, and that is the trap: it looks free. Sweep effort on your own evals before the swap, because the only independent measurement captured here covers one review harness on one workload 3. Budget for an input-token delta rather than parity, the way CodeRabbit measured it on its own workload 3, and treat every effort value inherited from Opus 4.8 as unset. Fast mode is a separate lever, latency: Anthropic says it runs around 2.5 times the default speed at twice Opus 5’s base price 1. Nothing is deprecated and no deadline is pushing you, so the useful move this week is that sweep.

Sources