Home › Glossary

What is reasoning effort (in plain English)?

Plain-English definition · Updated 2026-11-09. Every vendor claim below comes from that vendor's own documentation or launch page — OpenAI's GPT-5 developer guide and Help Center, Anthropic's Haiku 5.5 announcement and prompting docs, and Google's Gemini thinking documentation. Level menus change model by model; check the docs for the model you actually run.

AITop is an independent guide. Prices and features change fast — always check the vendor's page.

The 30-second answer

Reasoning effort is the setting that decides how many thinking tokens a model writes before it answers — which means it moves the bill and the wait, never the per-token price. In 2026 the three majors run three incompatible enums: OpenAI's GPT-5.2 accepts none / low / medium / high / xhigh (the pro tier drops none and low), Anthropic's output_config.effort takes low / medium / high / xhigh / max, and Google's Gemini 3+ retired the numeric thinking_budget in favor of MINIMAL / LOW / MEDIUM / HIGH string tiers. All three converged on the same default — medium. One trap is common to all of them: Anthropic's docs state that changing effort between requests invalidates your prompt cache.

What it actually means

Reasoning effort is a discrete tier that controls how much internal reasoning a model does before producing its visible answer. OpenAI's Help Center is explicit about the mechanism: "reasoning.effort controls how many reasoning tokens are generated before producing an answer" — and for GPT-5.2 the menu is none, low, medium, high, and xhigh, while gpt-5.2-pro only supports medium, high, and xhigh. Setting effort to none, OpenAI says, makes the model "behave like a non-reasoning model for latency-sensitive use cases." The original GPT-5 (August 2025) launched with a fourth value, minimal, added "to return an answer quickly"; the current 5.2 menu replaces it with none.

Anthropic's dial lives at output_config.effort and has five levels: low, medium, high, xhigh, max. Per Anthropic's own prompting guide, effort replaced the numeric thinking budget (budget_tokens) that older Claude models used — "there's no old setting to carry over" — and it covers every output token: text, tool calls, and thinking. Anthropic's Claude Code engineering guide puts it bluntly: "Effort is the main control for how long Claude thinks, and it replaces most of what people used to do with prompt wording." Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 default to medium; Claude Haiku 5.5 is the first Haiku-class model with an adjustable effort setting at all.

Google's Gemini 3 and later models use thinking_level with four values — MINIMAL, LOW, MEDIUM, HIGH — and Google's docs warn that the older numeric thinking_budget is deprecated: on Gemini 4 models it always returns a 400 error, on Gemini 3 it can fail intermittently, and sending both parameters in one request is an outright error. On Gemini 2.5 and earlier, the opposite holds: thinking_level isn't supported at all. The parallel with temperature's 2026 deprecation wave is not a coincidence — vendors are replacing continuous knobs with coarse, predictable tiers.

Why it matters when you're picking a tool

The dial is denominated in hidden tokens, so it is a bill dial, not a price dial. Effort levels do not change the per-token rate — they change how many reasoning tokens get written before the first visible word, all billed at the output rate. That makes effort one of the few settings that can swing an invoice by multiples without any line item looking different. Anthropic's own Haiku 5.5 launch page publishes accuracy-vs-cost charts per effort level with the cost axis on a log scale spanning $0.05 to $5 per attempt — more than two orders of magnitude on the same model. On the vendor's own evidence, the cheapest and most expensive effort level on one model differ by more than most models differ from each other.

Effort also changes behavior, not just length. OpenAI's developer guide notes that reasoning_effort controls "how readily it calls tools" — low-effort models reach for tools less. Anthropic's Haiku 5.5 guide warns the other direction: at low effort "the model is more likely to skip a search, stop early, or skip a check," which surfaces in agents as premature task hand-backs. If your agent feels lazy or chatty after a model update, effort is the first suspect — not the prompt.

Latency tracks effort almost linearly. Thinking tokens are output tokens: each one is generated by the same memory-bound decode step described in tokens per second. Low effort means fewer tokens to write before and after the first visible one, which is exactly why OpenAI offers none "for latency-sensitive use cases" and why Anthropic's docs report that thinking and replies "get much longer" at xhigh and max.

The 2026 reality check

"Medium" is the industry default — and that is worth noticing. OpenAI's reasoning_effort defaults to medium. Anthropic's Opus 5.5, Sonnet 5.5, and Haiku 5.5 default to medium. Google's thinking_level table lists MEDIUM as the default for most Flash models (with Pro-tier models defaulting HIGH and some Flash-Lite/image models defaulting MINIMAL). Three independent vendors converged on the same answer, which tells you where the cost-quality knee sits on current models. Treat medium as your eval baseline and only pay for deviations it earns.

"No thinking" means three different things. OpenAI has a true off switch: none on GPT-5.2. Anthropic's effort dial cannot turn thinking off at all — the separate thinking: disabled setting works at low, medium, and high, but Anthropic's docs state that requests with it at xhigh or max effort return a 400 error. Google's MINIMAL comes with a disclaimer in the official docs: it "does not guarantee that thinking is off" — the model may still reason minimally on complex tasks, and in multi-turn conversations MINIMAL requires thought signatures or the request fails with a 400. If your SLA depends on zero reasoning overhead, only one of the three vendors gives you a contract for it.

Switching effort has a hidden cache cost. Anthropic's docs: "Changing the top-level effort value between requests invalidates the prompt cache for the conversation's messages" — the effort value is part of the rendered prompt, so a medium-to-low flip mid-conversation re-bills the whole cache write (see prompt caching for the 1.25× write / 0.1× read arithmetic). Anthropic ships a per-message effort change as a beta (the mid-conversation-output-config-2026-07-01 header) precisely to let agents vary effort without breaking cache — and it requires adaptive thinking; with thinking off, it errors.

The top tiers have their own quirks. Anthropic documents that at xhigh in multi-turn chats the model "sometimes writes its whole answer in its thinking and ends the turn with no visible text" — parse for empty replies. And prompting does not substitute for the dial: in Anthropic's testing, telling the model to answer directly "didn't stop it from thinking." Meanwhile the level menus are deliberately uneven — gpt-5.2-pro trims none and low, Gemini 3.1 Pro tops out at HIGH with no MINIMAL — so a configuration that works on one model can 400 on its own pro sibling.

Quick checklist

Where you'll hit it

The effort dial sits behind three purchases on this site: the best AI chatbots ranking (a chatbot that suddenly "got slower" is often running a higher default effort tier), the ChatGPT vs Claude vs Gemini comparison (the same question at medium effort can cost measurably different amounts across the three, because each vendor's default tier and thinking behavior differ), and reasoning tokens (the invisible line on the bill that effort scales up and down). Related terms: prompt caching is the cost that breaks when effort flips, temperature is the sampling knob being deprecated in the same 2026 consolidation, and tokens per second is why more thinking tokens also means more waiting. Full list on the AI tools glossary.

The bottom line

Reasoning effort is the 2026 control surface that replaced prompt-hacking and numeric thinking budgets: OpenAI's GPT-5.2 takes none/low/medium/high/xhigh, Anthropic's effort takes low/medium/high/xhigh/max in output_config, and Gemini 3+ swapped thinking_budget for MINIMAL/LOW/MEDIUM/HIGH — three incompatible enums with one shared default, medium. Because the dial is denominated in output-billed thinking tokens, it moves cost and latency by multiples without touching the price list, and it even changes tool-calling behavior. Set it deliberately, eval two or three tiers, hold it constant within a conversation to protect your prompt cache, and never assume the top tier of one vendor maps to another's.