Every major AI model now has a setting that decides how long it thinks before it answers. OpenAI and xAI call it reasoning effort, Anthropic calls it effort, and Qwen combines a thinking mode with effort levels. In the apps the same dial appears as Instant to Extra High in ChatGPT, as an Effort menu in Claude and as Fast, Expert and Heavy in Grok.
The setting changes three things at once: answer quality, waiting time and cost. This guide explains how each vendor implements it as of October 2026, what the measured differences are and which level fits which task.
In short:
- Effort controls how many tokens a model spends on reasoning, tool calls and the answer, and reasoning tokens are billed as output even when you never see them.
- Returns diminish quickly. On GPT-6.1 Sol, going from low to medium adds six points on the Artificial Analysis Intelligence Index, the three steps above medium add four points together, and max costs about five and a half times as much per task as low.
- Medium is the right starting point for most work. Use low or off for simple high-volume tasks, and keep xhigh or max for long agentic runs where your own tests show a measurable gain.
What are effort levels in AI?
Effort levels tell a reasoning model how much work to put into a response. Before it answers, a reasoning model generates internal reasoning tokens in which it breaks the problem down, tries approaches and checks its own work. The effort level steers how many of those tokens it spends.
The idea behind it is test-time compute: instead of using a bigger model, you let the same model spend more computation per question. OpenAI’s reasoning guide describes the trade-off, with lower effort favouring speed and fewer tokens and higher effort letting the model think more completely. Effort is guidance rather than a hard cap, so at a low level a model still thinks on a genuinely hard problem, only less than it would at a high level.
What changes when you raise the effort level?
Raising the effort level changes quality, waiting time and price together, but none of the three moves in proportion to the others.
Quality: real gains that flatten out
Higher effort improves results on hard tasks, but each extra step buys less. The independent benchmarker Artificial Analysis tests the same model at every setting, which isolates the effect of effort, and its results for GPT-6.1 Sol show the pattern clearly.
| GPT-6.1 Sol setting | Intelligence Index score | Cost per benchmark task | Change from previous setting |
|---|---|---|---|
| Low | 42 | $0.13 | Starting point |
| Medium | 48 | $0.21 | +6 points for +$0.08 |
| High | 50 | $0.32 | +2 points for +$0.11 |
| Xhigh | 51 | $0.39 | +1 point for +$0.07 |
| Max | 52 | $0.72 | +1 point for +$0.33 |
The figures were read on 8 October 2026, and the index combines ten evaluations ranging from agentic coding to long-context reasoning. Going from low to medium adds six points for eight cents more per task. The three steps from medium to max add four points together and more than triple the cost, and the last step alone almost doubles the cost for a single point.
Speed: the model does not type slower, it types more
Higher effort is slower because the model generates more tokens, not because it generates them more slowly. On the same Artificial Analysis page, GPT-6.1 Sol’s output speed stays between 49 and 56 tokens per second across all five settings. Alibaba’s QwenCloud documentation shows the effect from the other side: on qwen3.7-plus, reasoning made up more than 60 percent of output tokens, and switching thinking off cut total latency by 60 to 75 percent.
Cost: you pay for thinking you never see
Reasoning tokens are billed as output tokens, which cost three to five times the input price for the models in this article. Our explainer on tokens in AI covers how token pricing works.
A worked example shows what that means. Take a request with 3,000 input tokens and a 500-token answer at $2 per million input tokens and $10 per million output tokens, the list price of both Claude Sonnet 5.5 and GPT-6.1 Sol. The reasoning token counts below are illustrative, as OpenAI notes that a single request can use anywhere from a few hundred to tens of thousands of them.
| Reasoning tokens per request | Cost per request | Cost per 10,000 requests |
|---|---|---|
| 0 (thinking off) | $0.011 | $110 |
| 2,000 | $0.031 | $310 |
| 15,000 | $0.161 | $1,610 |
The price per token is the same in every row, yet the bill grows almost fifteen times. Benchmarks confirm the effect at scale: Artificial Analysis measured Claude Sonnet 5.5 using about 193,000 output tokens per task at max effort, roughly seven times what GPT-6 Astra used at its own max setting.
How do effort levels work in ChatGPT and the OpenAI API?
ChatGPT uses thinking levels in the model picker, while the OpenAI API uses the reasoning.effort parameter.
Thinking levels in the ChatGPT app
Since GPT-6 began rolling out in October 2026, the Thinking slider offers Instant, Medium, High, Extra High and Pro. According to OpenAI’s help center, GPT-6 Sol powers the first four levels and Pro uses GPT-6 Pro, which is powered by GPT-6 Astra. Plus includes Instant, Medium and High, while Extra High and Pro require a Pro, Business or Enterprise plan. Free and Go users get a single Think option that runs on GPT-5.6 Luna. On Plus and Pro, Instant no longer moves up a level by itself when a question turns out to be complex, so pick the level yourself.
Reasoning effort in the OpenAI API
In the Responses API, reasoning.effort can be none, minimal, low, medium, high, xhigh or max, but each model supports its own subset. GPT-6 Astra does not accept none, and GPT-6.1 Sol supports neither none nor minimal and defaults to medium. A separate pro mode lets GPT-5.6 and GPT-6 models do more work on difficult tasks, independently of the effort level. On OpenAI’s price list, GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, GPT-6.1 Sol $2 and $10, and GPT-6 Luna $0.10 and $0.50.
How do effort levels work in Claude?
Claude offers five effort levels on its newest models: low, medium, high, xhigh and max.
Effort in the Claude apps
In the Claude apps, effort sits in the model menu next to the send button, and each model’s recommended level is marked as Default. Anthropic’s help center notes that higher effort takes longer and uses your plan’s limits faster, and it advises the defaults for everyday tasks and Max only for the most difficult, correctness-critical work.
The effort parameter in the Claude API
In the API, effort is set with output_config.effort. According to Anthropic’s documentation, most Claude models default to high, but Claude Opus 5.5 and Claude Haiku 5.5 default to medium, so a request that omits effort runs one level lower on Opus 5.5 than on Opus 5. Anthropic positions low for fast, cheap work such as subagents and xhigh for agentic and coding tasks that run longer than 30 minutes.
Effort applies to all output tokens, including tool calls, so a low-effort agent also makes fewer and shorter tool calls. On Claude Opus 5.5 adaptive thinking is always on, which makes effort the main lever for both depth and cost. Anthropic’s price list puts Opus 5.5 at $4 per million input tokens and $20 per million output tokens, Sonnet 5.5 at $2 and $10, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens.
How do effort levels work in Qwen?
Qwen offers the most granular control of the four, because Alibaba combines an on/off switch, a token budget and named levels, and because many Qwen models also run on your own hardware.
According to the QwenCloud documentation, hybrid Qwen models switch thinking on or off per request with enable_thinking, while thinking-only models always reason. The thinking_budget parameter caps the thinking tokens, after which the model answers immediately. Newer models add reasoning_effort, which on qwen3.8-max offers low, medium and xhigh. The default there is xhigh, which makes the most expensive setting the starting point, and the model rejects requests that combine reasoning_effort with thinking_budget. On the open-source Qwen3 hybrid models, adding /no_think to a message skips thinking for that turn.
QwenCloud lists qwen3.8-max at $2 per million input tokens and $6 per million output tokens, and some hybrid models charge a different rate in thinking mode. With open-weight models such as Qwen3.8-27B, which also defaults to xhigh, you pay for effort in GPU time and latency instead of per token, so the effort level becomes part of sizing your own hardware.
How do effort levels work in Grok?
Grok always reasons on its current models, so you choose how much it thinks but not whether it thinks.
xAI’s reasoning documentation lists low, medium, high and xhigh for grok-4.7 and grok-4.6, with high as the default, while grok-4.5 stops at high. On grok-4.20-multi-agent the same parameter sets how many agents work on a request, four or sixteen, rather than how long one model reasons. Grok 4.7 costs $2 per million input tokens and $6 per million output tokens.
The Grok apps use a mode picker instead. Fast answers quickly, Expert thinks hard, Auto chooses between the two, and Heavy puts a team of agents on the question for SuperGrok Heavy subscribers. Our complete guide to Grok AI covers the product in more depth.
Effort levels compared: ChatGPT, Claude, Qwen and Grok
Each vendor offers a fast setting, a balanced default and one or two levels for the hardest work, as the table below shows.
| Vendor | API parameter and levels | Default | In the app |
|---|---|---|---|
| OpenAI (ChatGPT) | reasoning.effort: none, minimal, low, medium, high, xhigh, max (subset per model) | Medium on GPT-6.1 Sol | Instant, Medium, High, Extra High, Pro |
| Anthropic (Claude) | output_config.effort: low, medium, high, xhigh, max | High; medium on Opus 5.5 and Haiku 5.5 | Effort menu plus a separate thinking toggle |
| Alibaba (Qwen) | enable_thinking, thinking_budget, reasoning_effort: low, medium, xhigh | Xhigh on qwen3.8-max | Thinking switch in Qwen Code |
| xAI (Grok) | reasoning_effort: low, medium, high, xhigh | High | Auto, Fast, Expert, Heavy |
Which effort level should you use for which task?
Start from the task rather than the model, because the right level depends on how much reasoning the work needs and what a wrong answer costs. The table combines the vendors’ own recommendations into one ladder.
| Task | Start at | Move up when | ChatGPT / Claude / Grok app |
|---|---|---|---|
| Classification, routing, extraction, simple lookups, voice | Off or low | The input is ambiguous or mistakes are costly | Instant / Low / Fast |
| Chat, customer support replies, short rewrites, subagents | Low | Answers skip steps or checks | Instant / Low / Auto |
| Drafting, summaries, data analysis, spreadsheets, everyday coding | Medium | The output is correct but shallow, or the task has many steps | Medium / Medium / Auto |
| Complex debugging, multi-step planning, contract or financial analysis, research synthesis | High | Quality still varies between runs on your test set | High / High / Expert |
| Long-running agents, deep research, security and code review, the hardest maths and coding | Xhigh or max | Only when your evaluations show a measurable gain over high | Extra High or Pro / Max / Heavy |
For high-volume tasks, start one level lower than you think you need and test whether quality holds, because the savings multiply with every request. When a mistake is expensive, such as a figure in a board report, start one level higher and still check the result.
Is a higher effort level always better?
A higher effort level is not always better. More thinking helps on problems that need reasoning, but it can hurt on tasks that do not. Researchers working with Anthropic built tasks in which longer reasoning lowered accuracy, as models got distracted by irrelevant details, latched onto spurious patterns or lost track of constraints. Anthropic’s guidance for Claude Opus 4.7 adds that max brings significant cost for small gains on most workloads and can cause overthinking on structured-output tasks. Effort is a tuning knob rather than a fix for a weak prompt, which is why OpenAI’s reasoning guide puts a clear goal, constraints and output format first.
Should you raise the effort level or switch to a bigger model?
Raising effort and switching models are two routes to better answers, and the cheaper route differs per case. Artificial Analysis found that Claude Sonnet 5.5 at low, medium and high effort sits behind GPT-6 Sol at high, xhigh and max, with Sol scoring higher while using fewer output tokens. Compare model and effort together rather than one at a time. In production this usually leads to routing, where a cheap model at low effort handles simple requests and only hard cases go further, as our article on LLM cost optimization explains.
Which mistakes make effort settings expensive?
Most surprises with reasoning models come from five avoidable mistakes.
- Teams accept the default without checking it, although defaults differ per vendor and even per model version, as the comparison table shows.
- The output limit is set too tight. It includes reasoning tokens, so a response can stop before any visible text appears while you still pay for the input and the reasoning, and OpenAI recommends reserving at least 25,000 tokens to start.
- Effort is changed in the middle of a conversation. On Claude a new top-level effort value invalidates the prompt cache, so use the per-message effort controls on Claude or the mid-conversation configuration update on GPT-6 models instead.
- Settings are carried over to a new model version. Anthropic notes that the levels of Claude Sonnet 5.5 are recalibrated, so the same level does not produce the same amount of thinking as on Sonnet 5.
- Effort is used to control length. It changes how much the model works, not reliably how much it writes, which Anthropic states explicitly for Claude Opus 5, so ask for the length you want in the prompt.
How do you find the right effort level for your own workload?
The reliable way is a small effort sweep: run the same real examples at two or three levels and pick the lowest level that still meets your quality bar. OpenAI and Anthropic both recommend testing on your own evaluations, because the effect of effort differs per task.
- Collect 20 to 50 real examples of the task, including a few hard cases, and define what a correct answer looks like.
- Run them at the default level and at one level below and above it, with the same prompt and model.
- Compare the share of correct answers, the output tokens per request and the time to a complete answer, using the usage data in each API response.
- Choose the lowest level whose quality you accept, and repeat the comparison when the model, the prompt or the task changes.
In a typical application the same model ends up at low for intake and classification, at medium for drafting and at high for the few steps where a mistake is costly. If you want help setting up evaluations, routing or cost control for your own use case, DataNorth’s AI consultancy team can help.
Frequently asked questions (FAQ)
What is the difference between effort and extended thinking?
Effort sets how much work a model puts into the whole response, while extended thinking is the separate reasoning step before the answer. In the Claude apps they are two settings: the effort level for thoroughness and a thinking toggle for whether reasoning appears first. On models with adaptive thinking, effort also decides how often and how deeply the model thinks.
Can you turn reasoning off completely?
You can turn reasoning off on some models, but not on all of them. OpenAI models that support the none setting, Claude models that allow thinking to be disabled and Qwen’s hybrid models can answer without reasoning. GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and the current Grok models always reason, so their lowest level is the fastest option.
Does asking a model to “think harder” change the effort level?
In ChatGPT it does not: OpenAI states that typing “think harder” or “think deeply” does not switch the thinking level, so select the level in the model picker. In the API only the effort parameter changes the level, although Anthropic suggests adding a line such as “this task involves multistep reasoning” when effort has to stay low for latency reasons.
Which effort level is best for coding?
Medium is a sound starting point for everyday coding, with high or xhigh for complex debugging and long agentic sessions. Anthropic recommends xhigh as the starting point for coding on Claude Opus 4.7 and 4.8, but medium for well-specified agentic coding on Claude Sonnet 5.5, so test two levels on your own codebase.
