Claude Opus 5 Effort Toggle: How to Trade Cost for Capability
Deep dive into Claude Opus 5's low/medium/high effort settings: how they work, what they cost, and how to route requests for the best cost-quality ratio.
TL;DR
Claude Opus 5 introduces a per-request effort setting — low, medium, or high — that controls how much reasoning the model invests before answering. Low effort is faster and cheaper for mechanical tasks; high effort buys longer reasoning for the hardest problems. Teams that route requests by difficulty instead of running everything at one setting typically cut spend substantially while improving results where it matters.
What the Effort Toggle Actually Does
Extended thinking is on by default in Opus 5. The effort setting governs how much of it the model uses: how long it deliberates, how many alternatives it explores, and how aggressively it verifies its own work before responding. Because output tokens cost $25 per million, thinking depth translates directly into cost — the toggle is effectively a price dial that also moves quality and latency.
The Three Settings
Low. Minimal deliberation. Best for: formatting and extraction, simple edits, classification, summarization of clean inputs, high-volume pipeline steps. Quality on these tasks is usually indistinguishable from high effort — you're just paying less and waiting less. Low-effort Opus 5 competes on cost with much smaller models while retaining Opus judgment. Medium. The sensible default. Balanced reasoning for everyday coding, document analysis, and agent steps. If you set one value globally, set this one. High. Maximum deliberation. Best for: gnarly debugging, architectural decisions, multi-constraint planning, security review, anything where a wrong answer is expensive to discover later. Expect more latency and more output tokens — worth it exactly when iteration cost exceeds token cost.Routing Patterns
Static routing by endpoint. Map each pipeline stage to a fixed effort level: extraction jobs run low, agent planning runs high, everything else runs medium. Simple, predictable, and captures most of the savings. Escalation on failure. Start at low or medium; if validation fails (tests don't pass, output doesn't parse), retry at a higher effort. This pays the high-effort premium only on requests that proved hard. Two-pass drafting. Draft at low effort, then have a high-effort pass review and correct. For long documents this often beats a single high-effort generation on both cost and quality.Interaction with Fast Mode
Effort and fast mode are independent dials: effort controls thinking depth, fast mode ($10/$50, ~2.5× speed) controls serving speed. A latency-sensitive but hard task can run high effort + fast mode; a batch job can run low effort + standard speed for minimum cost.
Measuring the Impact
Track three numbers per route: cost per successful task, retry rate, and p95 latency. The common finding after a week of data: most traffic runs fine on low or medium, and concentrating high effort on the genuinely hard 10-20% of requests improves aggregate quality while lowering the bill.
Conclusion
The effort toggle turns model selection from a one-time decision into a per-request one. Route mechanical work to low, keep medium as the default, reserve high for problems where being right the first time is worth paying for — and let escalation-on-failure catch the misclassifications.