Claude Opus 5 Pricing and Reasoning Effort: The Settings Guide Nobody Wrote
Claude Opus 5 pricing, reasoning effort levels, and fallback behavior explained: why max thinking wastes money and what settings actually perform best.
What does Claude Opus 5 cost, and what reasoning level should you actually use?
Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, the same rate as its predecessor Opus 4.8. On reasoning effort, more thinking is not automatically better: benchmark data from Anthropic itself shows performance on coding tasks peaking at medium effort and getting worse, not better, once you push past it. For most day-to-day use in Claude Code or similar workflows, medium is the effort level to start with, with high reserved for cases where you’ve tested and confirmed it helps.
TL;DR
- Pricing stayed flat: Opus 5 costs $5/million input tokens and $25/million output tokens, identical to Opus 4.8, so the model got smarter without getting more expensive per token.
- Reasoning effort has a ceiling: Anthropic’s own data shows Opus 5 scoring best on Frontier coding benchmarks at medium effort (around 53%), then declining in accuracy while costing more as you push to high or max.
- Max thinking often burns money for nothing: cranking reasoning effort to maximum can mean paying more tokens for a worse answer, because the model overthinks problems that had an obvious solution path.
- Independent evaluators lean toward high, not medium: Vals AI found high reasoning effort produced its best results in their own testing, so the right setting depends on the task type and isn’t identical across benchmarks.
- Cost-effectiveness beats raw intelligence in most work benchmarks: across agentic coding, computer use, and business workflow tests, Opus 5 matched or beat larger models like Claude Fable 5 at roughly half the cost.
- Special cases need special settings: Anthropic’s IMO 2026 proof-solving run used adaptive thinking set to max, with failed attempts automatically resampled at lower effort, showing that some tasks do reward pushing reasoning higher.
- The Epoch Capabilities Index puts Opus 5 behind Fable 5 and GPT-5.6, a reminder that on pure aggregate intelligence scoring, Opus 5 isn’t the top model, it’s the best value model.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Why doesn’t more reasoning effort mean a better answer?
Reasoning models work by generating an internal chain of thought before producing a final response, and the general assumption has been that longer chains produce better answers, at the cost of more tokens and more money. Opus 5 breaks that pattern in a specific, measurable way.
On Frontier coding benchmarks, Anthropic’s data shows Opus 5’s accuracy peaking at medium reasoning effort. Pushing the setting to high or max doesn’t just cost more, it makes the model perform worse. The likely explanation is overthinking: when a problem has an obvious solution, extended reasoning gives the model room to second-guess itself, wander into unnecessary tangents, or talk itself out of a correct first instinct. This isn’t unique to Opus 5. Smaller reasoning models in general tend to show diminishing or negative returns once effort is cranked past a certain point, which is why running something like a lightweight model on maximum thinking is often a waste of budget rather than a quality upgrade.
The practical rule: treat “max effort” as a setting for genuinely hard, open-ended problems, not a default. For routine coding, agentic workflows, and general queries, medium effort is the more defensible starting point.
Is high or medium the better default for Opus 5?
The honest answer is that it depends on who’s measuring and what they’re measuring. Two credible sources point in different directions.
Vals AI, an independent benchmarking group, found that high reasoning effort produced the strongest results across their test suite when evaluating Opus 5. Anthropic’s own published data, by contrast, shows accuracy on coding-specific benchmarks declining once effort moves past medium. Both can be true at once: the ideal reasoning level appears to be task-dependent. Coding and agentic work, where there’s often a clean, verifiable answer, seem to reward stopping at medium. Broader reasoning or open-ended analysis tasks, the kind Vals AI tests across professional workflows, may benefit from the extra deliberation that high effort provides.
The practical takeaway for anyone running Opus 5 in production: don’t assume one setting fits every workload. Start at medium for coding and agentic tasks. Test high on tasks that involve ambiguity, multi-step planning, or synthesis across large amounts of context. Reserve max effort for genuinely exceptional cases, since it’s the most expensive setting and the one most likely to underperform on tasks that don’t need it.
When does it make sense to push reasoning to max?
There is at least one clear example where maximum reasoning effort was the right call: Anthropic’s own evaluation of Opus 5 against the 2026 International Mathematical Olympiad problem set. The model was asked to produce rigorous, self-contained proofs without any agent harness or external tools, working within a 256,000 token output limit. Reasoning effort was set to adaptive max, meaning the model could use as much internal deliberation as the problem demanded. When an attempt exhausted the output limit without finishing, it was resampled at a lower thinking effort rather than simply failing.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
The result was a perfect score, 42 out of 42, which corresponds to gold medal level performance on that competition, well above the human gold medal threshold. That’s a case where the task is genuinely hard, has no shortcut, and rewards exhaustive reasoning. It’s the opposite profile of routine coding tasks, where the answer is usually reachable quickly and extended thinking just adds noise and cost.
The lesson generalizes: match reasoning effort to the actual difficulty and open-endedness of the task, not to a blanket assumption that more thinking is always safer.
Is Claude Opus 5 worth using over Claude Fable 5?
On a narrow “which model is smarter” basis, the answer is mixed. On the Epoch Capabilities Index, a meta-benchmark that aggregates dozens of individual tests using item response theory (a statistical method borrowed from educational testing) to avoid saturation as models approach ceiling performance on individual benchmarks, Opus 5 sits behind both Fable 5 and GPT-5.6. So if raw aggregate intelligence is the only metric that matters, Opus 5 isn’t the top model available.
But cost changes the calculation. Across a range of work-focused benchmarks, agentic business workflows, computer use tasks, and agentic coding, Opus 5 matched or outperformed Fable 5 while costing roughly half as much. On the Apex Agents benchmark, which tests professional tasks like investment banking analysis, management consulting, and legal work using expert-graded rubrics, Opus 5 scored marginally higher than Fable 5, at half the price. On software-engineering-specific evaluations built with Cognition, Fable 5 held a slight edge, but not by a margin that changes the economics.
For most people running real workflows rather than chasing leaderboard position, the pricing gap matters more than a percentage point or two of accuracy. If your use case is agentic coding, business automation, or computer-use tasks, Opus 5’s cost-to-performance ratio makes it the more practical default. If you specifically need the highest ceiling on open-ended reasoning or novel problem types, it’s worth testing Fable 5 or a comparable frontier model against your specific workload before committing.
Frequently Asked Questions
What is Claude Opus 5’s pricing per million tokens?
Opus 5 costs $5 per million input tokens and $25 per million output tokens. That’s the same pricing as Opus 4.8, meaning capability improved without a price increase.
Should I set Claude Opus 5’s reasoning effort to max by default?
No. Anthropic’s own benchmark data shows accuracy on coding tasks peaking at medium effort and declining at higher settings, so setting reasoning to max by default often costs more while producing worse results.
Does the best reasoning level change depending on the task?
Yes. Independent testing from Vals AI found high effort produced the best overall results in their benchmark suite, while Anthropic’s coding-specific data favored medium. Task type matters more than a universal default.
Is Claude Opus 5 more intelligent than Claude Fable 5?
On aggregate intelligence measures like the Epoch Capabilities Index, Opus 5 ranks behind both Fable 5 and GPT-5.6. Its advantage shows up in cost-effectiveness on work-related benchmarks, not in topping every intelligence ranking.
When should I use maximum reasoning effort on Opus 5?
Reserve it for genuinely hard, open-ended problems without an obvious solution path, such as rigorous proof writing or multi-step problems where failed attempts can be resampled at lower effort. Routine coding and agentic tasks generally perform better, and cheaper, at medium effort.
