How reasoning effort changes LLM cost per task, and how to set it

Reasoning effort is the setting that tells a model how much to think before it answers, and the thinking is billed as output tokens. On Artificial Analysis’s leaderboard, as of 28 October 2026, the five models in our table cost 4.4 to 13.5 times as much per task at their highest setting as at their lowest. The wait for the first chunk of output grows from seconds to minutes. In all five, low to medium gave the biggest score gain, and xhigh to max was the most expensive step.

Medium is the documented default for four of the five, so on those four a request that sets nothing runs at medium.¹² Turning effort up on a cheaper model can also beat paying for a pricier one. GPT-6.1 Sol at max scores 55 on the index, the same as the more expensive GPT-6 Astra at high, for 39% of the cost per task, though it takes 251 seconds to start answering against 38.

Below are the API settings, the data and its method, and a way to choose a level for one feature. We haven’t run these benchmarks. Scores, costs and latencies are Artificial Analysis’s third-party measurements, and API details come from vendor documentation, all read on 28 October 2026.

What the effort setting changes in a request

Reasoning models write hidden tokens before the answer. OpenAI says these take up context window space and are billed as output tokens.³ Anthropic bills Claude’s full thinking as output even when the response shows only a summary of it or nothing, and Google bills Gemini’s thinking the same way.⁴⁵ Effort tells the model roughly how much of this to do.

The model still decides on each request. OpenAI says its models use fewer reasoning tokens on simpler tasks at any level.³ Anthropic says effort is a behavioural signal with no strict token budget, and that at lower levels Claude can skip thinking on easy requests.¹ Google describes Gemini’s default as dynamic thinking that adjusts to how complex the request is.⁵

More thinking means a longer wait. OpenAI says lower effort favours speed, and Google lists high as the level with the slowest first token on Gemini 4 Argon.³⁶

On Claude, effort applies to every output token, so it changes the text and the tool calls as well as the thinking. Anthropic says lower levels make fewer and terser tool calls, and that in long agent prompts Claude Haiku 5.5 at low is more likely to stop early or skip a search.¹ Level names don’t carry across models either. Anthropic recalibrated Claude Sonnet 5.5’s levels, so each one gives a different amount of thinking from the same level on Claude Sonnet 5.

OpenAI reasoning effort, Claude’s effort parameter and Gemini’s thinking level

The table summarises each vendor’s documentation as of 28 October 2026.

Claude’s older thinking budget, budget_tokens, is rejected from Claude 4.7 on, so effort is now the main cost control on Claude.⁷¹

Reasoning effort cost per task on Artificial Analysis

Artificial Analysis benchmarks each effort level as a separate entry. Each cell below shows the Intelligence Index score and the cost per task in US dollars, as of 28 October 2026.⁹

The ratios use Artificial Analysis’s unrounded costs. OpenAI released GPT-6.1 Terra on 26 October 2026.¹⁰ The label on the Opus 5.5 rows is covered in the methodology section. Claude Sonnet 5.5 carries the same label and spans a wider range than any model in the table, from 25 at low for $0.51 to 54 at max for $10.02, about 20 times the cost.

The next table gives the median seconds to the first chunk of the response. For these models that is the first answer token, so it includes the thinking. Artificial Analysis measures it on a separate prompt of about 10,000 input tokens, over the previous 72 hours.¹¹

At medium, all five start answering within about 15 seconds. At max, none starts within two minutes, and Claude Opus 5.5 takes about 11 minutes.

Google’s models have thinner coverage. Gemini 4 Argon appears only at high, scoring 50 for $7.68. Gemini 4 Carbon tops the index at 59 for $7.76, also only at high, and Artificial Analysis says it isn’t publicly available yet.¹² Gemini 3.8 Flash appears at three levels, but its low setting has no cost listed. Among open-weight models, only GLM-5.3 is scored at two settings, 23 at low for $1.57 and 32 at max for $5.01.⁹

Where extra effort buys points and where it mostly buys cost

In all five models in the table, low to medium gave the largest score gain of any step, from 6 points on Haiku 5.5 to 13 on Opus 5.5. The step from xhigh to max was the most expensive in all five, taking 43% to 63% of each model’s cost increase from low to max, and it added 1 to 3 points, or 6 on Terra.⁹

Some steps buy almost nothing on the index. GPT-6 Astra scores 55 at both high and xhigh, and xhigh costs 13% more. Grok 4.7 drops from 35 at high to 34 at xhigh, which costs 21% more, and Gemini 3.8 Flash goes from 30 at medium to 31 at high for 24% more, $7.77 a task against $6.26. On a composite of nine evaluations we’d treat a gap of a point or two as level and take the cheaper setting.

A matching score often costs less on a cheaper model at a higher setting. GPT-6.1 Sol at max equals GPT-6 Astra at high, 55 each, for $1.27 a task against $3.24. GPT-6.1 Terra at max equals Astra at medium, 51 each, for $0.96 against $2.73. Claude Haiku 5.5 at max scores 36 for $0.41, against 35 for $0.98 from Claude Opus 5.5 at low. Artificial Analysis notes that Sol at max comes within 4 points of Opus 5.5 at max for about an eighth of the cost.¹³

The cheaper route is slower. Sol at max takes 251 seconds to its first chunk against 38 for Astra at high, Terra at max 138 against 3.9, and Haiku 5.5 at max 343 against 5.2 for Opus 5.5 at low. Waits of two to six minutes rule these pairings out for most features a user watches.

Most of the cost per task is input, and it rises with effort

Artificial Analysis also splits cost per task by token type on its model pages. At all 25 settings in the first table, input was the largest part, 47% to 73%, most of it cache reads and writes. Reasoning was 7% to 39%.¹⁴

Input cost grew with effort as well, about 4 times on Astra, 6 times on Sol and 10 times on Haiku 5.5 from low to max. Artificial Analysis doesn’t say why, but two documented behaviours would push it up in an agent loop. Anthropic says higher effort tends to make more tool calls, and each tool call means another request carrying the conversation so far. On Claude, thinking blocks from earlier turns that stay in context are also billed as input on later requests.¹⁴ A single short request has no history to resend, so reasoning should be a larger share of its bill than these figures show.

How Artificial Analysis measures cost per task

Artificial Analysis defines cost per task as the weighted average cost, in US dollars, of one Intelligence Index task. It prices the input, cached and output tokens each model used, weights each evaluation as the index does and divides by the number of tasks.¹⁵ Token counts come from each provider’s API, cache use is costed at each endpoint’s typical hit rate, and prices are the maker’s own API rates or the median across providers. Reasoning models get the largest output limit their makers allow.¹⁶¹⁵

These figures come from Intelligence Index v5.0, announced on 27 October 2026. It has nine evaluations in four equally weighted categories, and 60% of its weight rests on private test sets or held-out answers. Artificial Analysis says four agentic evaluations carry 55% of the weight but, for the median model, about 90% of the output tokens, and that effort matters more than under v4.3. Sonnet 5.5’s spread from low to max grew from 20 to 29 points, and Sol’s from 10 to 20.¹³ Figures from v4.3.2, including those in our October comparison of LLM API prices and benchmarks, aren’t comparable.

Every Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 row is labelled “with fallback”, or “Default Fallback” on the model pages, and we found no explanation in Artificial Analysis’s methodology. The name matches Anthropic’s server-side fallback, a beta option that retries a request declined by a safety classifier on a model Anthropic recommends for that refusal category. Claude Haiku 5.5 has no server-side fallback, and its rows carry no label.¹⁷ We couldn’t confirm what the setting changes in these runs or how many tasks it touched, so no conclusion here rests on those rows alone.

An index score is an average over many kinds of task, mostly long agent runs, and a single feature’s tasks may need far more or far less reasoning than that mix.

How to choose an effort level for one feature

The rule we’d follow is to use the cheapest level that passes your own evaluation and fits the feature’s latency budget, set per feature. A ticket classifier, a nightly report and a coding agent in one product are unlikely to want the same level. Anthropic’s suggestions for Claude Haiku 5.5 already vary by job, with low for chat and simple high-volume requests and high for knowledge work and longer agent tasks.¹

The vendors advise testing too. Google tells Gemini 4 Argon users to evaluate all three levels and take the lowest that meets their bar. Anthropic tells Opus 5.5 users to run an effort sweep on their own evals, and OpenAI says to use xhigh only when evals show a clear benefit.⁶¹³ A method:

1. Build an eval set from the feature’s real traffic, with checks in code where possible. Our one-week routine for model upgrades covers how.
2. Run it at the default and one level either side, usually low, medium and high, and repeat each run.
3. Record for each task whether it passed, its input, output and reasoning tokens, its cost, and its p50 and p95 time to the first visible token.
4. Divide cost by pass rate to get cost per passing task, then pick the cheapest level that clears your quality bar and latency budget.
5. Keep the level in config next to the model ID, and run the sweep again when the model, its default or your prompt changes.

A worked example

These figures are assumptions for the arithmetic, and we didn’t measure them. A ticket-routing feature runs on GPT-6.1 Sol, at $2 per million input tokens and $10 per million output tokens as of 28 October 2026.¹⁸ It handles 20,000 requests a day, each with 4,000 uncached input tokens and a 300-token answer. Suppose the eval shows 500 reasoning tokens a request at low with 91% passing, and 3,500 at high with 95% passing.

High adds $0.030 a request and cuts failures from 9% to 5%. It pays for itself only if a failure costs more than $0.030 ÷ 0.04 = $0.75 to put right, in a support agent’s time for example, and only if its p95 latency fits the feature.

Interactions that change the bill

Caching discounts input only. Reasoning tokens are output, so a cache hit never makes thinking cheaper. Changing effort can also break the cache. Anthropic renders the effort level into the prompt, so a new top-level value invalidates cached prefixes. Its per-message effort change, in beta on the Claude API for all four current models, keeps them.⁴¹ OpenAI’s equivalent is a configuration_update item.³

Batch pricing halves reasoning too. Anthropic’s Batch API takes 50% off input and output, OpenAI prices Batch and Flex at half its standard rates, and Gemini 4 Argon’s batch output price is half its standard one.¹⁹²⁸ Evals, backfills and nightly reports can run at any level for half the price.

Output limits truncate long thinking. All three vendors count reasoning against the output limit. An OpenAI response that hits max_output_tokens comes back incomplete, possibly before any visible text, and you still pay for the reasoning. OpenAI suggests reserving at least 25,000 tokens while you experiment.³ Anthropic’s max_tokens caps thinking and text together for each request, so it doesn’t bound a whole agent turn.⁴ Gemini stops with an incomplete status and still bills the thinking, and synchronous Gemini 4 responses stop at 65,536 tokens.⁵

Latency budgets. Google says hard Gemini 4 Argon prompts can take several minutes and recommends client timeouts of at least 15 minutes.⁶ OpenAI suggests asking the model for a short preamble when the time to the first visible text matters.³ The latency table shows medians, so measure p95 at your chosen level.

Defaults change between versions. Claude Opus 5.5 defaults to medium where Claude Opus 5 defaulted to high, so a request that sends no effort runs a level lower on the newer model, and Claude Sonnet 5.5’s levels were recalibrated.¹ Set the level explicitly and include an effort sweep in every model upgrade.

A checklist for setting reasoning effort

1. Set effort explicitly for each feature, in config next to the model ID.
2. Move up a level only when the failures it prevents cost more than the extra spend, and the latency still fits.
3. Log reasoning tokens on every call: reasoning_tokens on OpenAI, thinking_tokens on Claude and total_thought_tokens on Gemini.
4. Leave room for thinking in the output limit, and alert on truncated responses.
5. Keep effort fixed within a cached conversation, or use the per-message changes that keep the cache.
6. Send high-effort work that can wait through batch pricing.
7. Run the sweep again when the model, its default or your prompt changes.

Frequently asked questions