Reasoning effort settings - when high vs low pays off

Most modern reasoning models let you turn a dial: how hard should the model "think" before it answers? OpenAI calls it reasoning_effort, Anthropic exposes a thinking budget, and other vendors ship similar knobs. Cranking it to high feels safe — but it costs more tokens, adds latency, and on plenty of tasks it changes nothing. This guide is about spending that budget where it actually moves the needle.

What the setting actually controls

Reasoning effort governs how many internal "thinking" tokens a model spends before producing its visible answer. More effort means more intermediate reasoning: the model explores alternatives, checks its own work, and backtracks. That extra reasoning is billed as output tokens, so high effort is not free — you pay for tokens you never read, and you wait longer for the first useful response.

The trade-off is real but not linear. On a hard multi-step problem, the jump from low to high can be the difference between a wrong answer and a correct one. On a simple lookup or a formatting task, it just burns tokens.

When high effort pays off

Reach for high (or medium-high) when the task has a verifiable correct answer and several dependent steps:

• Multi-step math, logic, or proofs — anything where one early mistake cascades.
• Complex code generation and debugging — tracing state across functions, reasoning about edge cases, refactoring with constraints.
• Planning and decomposition — agent workflows that must break a goal into ordered sub-tasks.
• Structured extraction from messy input — reconciling conflicting fields, applying rules with exceptions.
• Ambiguous requirements — where the model must weigh options before committing.

The signal: if a human expert would need scratch paper, the model probably benefits from a bigger thinking budget.

When low effort wins

Low effort is the right default more often than people assume:

• Classification, tagging, sentiment, routing — short answers from clear inputs.
• Summarization and rewriting — style transforms, not deductions.
• Lookups, formatting, simple transformations — JSON cleanup, casing, templating.
• High-volume, latency-sensitive endpoints — chat UIs, autocomplete, anything user-facing where speed is the feature.

On these, high effort adds cost and lag without improving the result, and occasionally makes it worse by over-thinking a trivial prompt.

A quick decision table

Test it, don't guess

The honest answer is that the right setting is task-specific, so measure it. Run a sample of your real prompts at low and high, then compare accuracy, latency, and token cost. The cheapest way to do that is on a single endpoint where you can also swap models without touching auth or base URLs.

With AnyModel you get one OpenAI-compatible endpoint (https://anymodel.org/v1) and one API key that reaches GPT, Claude, Gemini, DeepSeek, GLM, Kimi, Qwen, and Grok. Switching model or effort is just a field change:

curl https://anymodel.org/v1/chat/completions \
  -H "Authorization: Bearer $ANYMODEL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5",
    "reasoning_effort": "low",
    "messages": [{"role": "user", "content": "Classify: \"refund my order\""}]
  }'

Flip reasoning_effort to "high", or change "model" to a different provider, and rerun — nothing else moves. That makes A/B testing effort levels and models genuinely cheap. Want to see which model is strongest per task before you tune effort? Our model comparison is a good starting point, and the blog has more hands-on guides.

A practical default

Start low. Promote to high only for the task types above, and only after a test shows it actually improves the outcome you care about. For sensitive prompts you'd rather not have stored, Ghost Mode offers opt-in zero-retention keys — we keep only a token counter, though the model provider still receives your prompt.

You get 1,000,000 free tokens on signup (6,000,000 total if you link Telegram), no credit card, then simple pay-per-token with no subscription. That's plenty to benchmark high vs low across a few models and find your own sweet spot.

Create your free AnyModel account and start testing reasoning effort today.