Claude effort vs adaptive thinking: what each one controls

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. Your batch service on Claude Opus 5 (claude-opus-5) never shows reasoning to end users. An engineer proposes thinking: {"type": "adaptive", "display": "omitted"} on every request, and you ask what the setting actually saves on the bill and on latency before it ships. How should you evaluate this proposal?

   Right. In the game that’s Renown for your Prompt Engineering, Structured Output and Retrieval building.

   Not that one: A is right. In the game your Prompt Engineering, Structured Output and Retrieval building would lose a little Renown, and this question would come back later.

   The answerA.Ship it for latency alone, because the thinking tokens are still billed in full; omitting them reduces the time to first text token, not cost.With `display: "omitted"Your answerB.Ship it for the bill alone, because the thinking tokens are billed at the cache read rate once the reasoning is suppressed, and the gain is in cost per turn.Imports a caching multiplier into a display setting that has nothing to do with the cache. Suppressing the returned reasoning does not move those tokens onto the cache read rate.Your answerC.Ship it for both gains, because the thinking tokens are never generated at all and the gain is in cost and in latency together.The reasoning still happens; only its text is withheld. The signature field still carries the encrypted full thinking so a later turn can continue from it.Your answerD.Ship it for the bill alone, because the thinking tokens are billed at the input rate and the gain is in the lower input price.Reclassifies thinking tokens as input. They are billed as output tokens whatever the display setting is, so the rate does not change.

   Show the full explanation and doc quotes

   ` the thinking blocks come back with an empty `thinking` field, but the model still reasons and the full thinking tokens are still billed as output. The documented benefit is faster time to first text token when streaming, because the server skips streaming thinking tokens and delivers only the signature.

   Written from platform.claude.com/docs/en/build-with-claude/thinking

2. You want Claude Sonnet 5 (claude-sonnet-5) to decide per request how deeply to reason, and you write output_config: {"effort": "adaptive"} on the request. The API rejects the value. What is the most likely root cause?

   Right. In the game that’s Renown for your Prompt Engineering, Structured Output and Retrieval building.

   Not that one: B is right. In the game your Prompt Engineering, Structured Output and Retrieval building would lose a little Renown, and this question would come back later.

   Your answerA.adaptive is valid only when thinking is omitted from the request entirely.adaptive never becomes a valid effort level. It is the value that thinking.type takes when Claude decides for itself when and how deeply to think.The answerB.adaptive is a thinking mode, not an effort level, so it is invalid for effort.The thinking parameter controls whether Claude reasons in thinking blocks and takes values such as adaptive, enabled and disabled. The effort parameter controls how much work Claude puts into the whole response and takes low, medium, high, xhigh or max.Your answerC.adaptive is valid for effort but only on models that also support xhigh.xhigh availability varies by model — Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Opus 4.8, Claude Opus 4.7 and Claude Sonnet 5 — but that has no bearing on adaptive, which is not an effort level on any model.Your answerD.adaptive is the default for effort and must never be sent explicitly.The API default for effort is high, and passing high explicitly behaves identically to omitting the parameter.

   Show the full explanation and doc quotes

   The effort documentation warns explicitly not to pass adaptive as an effort value.

   Written from platform.claude.com/docs/en/build-with-claude/effort

3. You maintain a shared prompting library used against several Claude models. One code path still sends thinking: {type: "enabled", budget_tokens: 8000}. On Claude Sonnet 4.6 (claude-sonnet-4-6) requests succeed, but the identical call against Claude Opus 5 (claude-opus-5) returns 400. What is the most likely root cause?

   Right. In the game that’s Renown for your Prompt Engineering, Structured Output and Retrieval building.

   Not that one: D is right. In the game your Prompt Engineering, Structured Output and Retrieval building would lose a little Renown, and this question would come back later.

   Your answerA.budget_tokens is rejected on Opus 5 because 8000 exceeds that model's per-request thinking ceiling, which Sonnet 4.6 sets higher.The rejection is categorical, not a bounds check. No value of budget_tokens is accepted on Claude Opus 5 because the parameter has been removed from that model's request surface entirely.Your answerB.budget_tokens is rejected on Opus 5 because the request omits the extended-thinking beta header that model requires and Sonnet 4.6 does not.Adaptive thinking on Claude Opus 5 is generally available and requires no beta header. Adding one would not revive a parameter the model no longer accepts.Your answerC.budget_tokens is rejected on Opus 5 because fixed budgets are valid only when max_tokens is set above them, a check Sonnet 4.6 skips.The rule that budget_tokens must be less than max_tokens — with a minimum of 1024 — applies to Sonnet 4.5, Haiku 4.5 and older models that still take fixed budgets. It has no bearing on the 400 from Opus 5.The answerD.budget_tokens is rejected on Opus 5 because the parameter has been removed from that model entirely, while Sonnet 4.6 still accepts it.Fixed thinking budgets were deprecated on Claude Opus 4.6 and Sonnet 4.6, where they remain functional as a transitional escape hatch, and were fully removed on Claude Fable 5, Opus 5, Opus 4.8, Opus 4.7 and Sonnet 5, where budget_tokens returns 400.

   Show the full explanation and doc quotes

   Adaptive thinking with output_config.effort is the replacement.

   Written from platform.claude.com/docs/en/build-with-claude/thinking