Ask Claude Sonnet 5.5 a hard question at maximum effort and it will write around 193,000 tokens of private reasoning per benchmark task before it hands over an answer. You never read most of that. You pay for every token of it at the output rate of $10 per million, which is how Artificial Analysis ended up paying $7.60 per task to test it.
Reasoning models got smarter over the last two years by talking to themselves, and that self-talk is now the biggest line on a lot of API bills. Every major provider gives you a dial to turn it down. Most teams leave it on the factory setting and never look. This instalment of our AI on a Budget series covers that dial: what it is called at each provider, what each notch costs, and how far down you can turn it before quality falls off.
How thinking became the expensive bit
OpenAI’s o1 in September 2024 set the template: the model reasons in hidden tokens, you get billed for them as output, and you see a count rather than the text. Two years on, every frontier lab works the same way, and the controls have multiplied. OpenAI’s reasoning.effort now runs from none through minimal, low, medium, high and xhigh to max, though no single model accepts all seven. Anthropic has retired its old fixed budget_tokens in favour of adaptive thinking plus an effort level. Google moved Gemini from thinking_budget to thinking_level.
The billing rule is identical everywhere. Thinking is charged as output, and output costs five times input on the current price cards: Claude Sonnet 5.5 and GPT-6.1 Sol are both $2 in and $10 out per million tokens, Claude Opus 5.5 is $4 and $20. Anthropic’s docs spell out the part people miss: you are billed for the full thinking process even if you ask for a summary, or for no thinking text at all.
What each notch buys you
Artificial Analysis ran GPT-6.1 Sol through its Intelligence Index at all five supported effort levels:
- Low: score 42, $0.13 per task, about 4,000 output tokens.
- Medium: score 48, $0.21 per task.
- High: score 50, $0.32 per task.
- Xhigh: score 51, $0.39 per task.
- Max: score 52, $0.72 per task, about 38,100 output tokens.
Medium captures six of the eight points between low and high for 62% more money. Going from medium to max adds four more points and multiplies cost by roughly 3.4. Output speed barely moves across the levels, so the extra cost and the extra waiting both come from the model writing more. The same shape shows up on GPT-6 Astra, where a task costs $0.82 at low and $3.26 at max.
Effort also interacts with model choice. At max, Sonnet 5.5 used about seven times the output tokens GPT-6 Astra did, and its cost per task came in about 50% above its predecessor Sonnet 5, at an unchanged sticker price. Price per token tells you very little. Price per finished task is the number that lands on your invoice.
The maths on your own traffic
Take a fairly ordinary request: a 3,000-token prompt and a 500-token answer, at $2 in and $10 out. With no thinking it costs $0.011. Add 2,000 thinking tokens and it costs $0.031. Let the model ramble to 15,000 thinking tokens and it costs $0.161. Over 10,000 requests a month that is $110, $310 or $1,610, and the answer the user reads is the same length in all three.
Now say half that traffic is simple work (classification, extraction, reformatting) currently running at the heavy setting. Dropping those 5,000 requests from 15,000 to 2,000 thinking tokens saves $650 a month on a $1,610 bill, roughly 40%, with no change of model and no code beyond one parameter.
The exact settings
- OpenAI: pass
reasoning: {effort: "low"}in the Responses API. Most current models default tomedium. GPT-6.1 Sol acceptslowand up only; GPT-6 Astra returns a 400 error if you sendnone. Cap spend withmax_output_tokens, which covers reasoning plus visible output. OpenAI suggests reserving at least 25,000 tokens while you experiment. - Anthropic: send
thinking: {type: "adaptive"}plusoutput_config: {effort: "medium"}. Levels arelow,medium,high,xhighandmax. The default ishighon most models andmediumon Opus 5.5 and Haiku 5.5. The oldbudget_tokensroute is deprecated, and Claude 4.7 and later reject it outright. Readusage.output_tokens_details.thinking_tokensto see what you were billed for. - Google: set
thinking_level. Gemini 3.1 Pro takeslow,mediumorhighand defaults tohigh; the Flash models addminimal. Sending boththinking_leveland the legacythinking_budgetin one request gets you a 400.
Where turning it down goes wrong
First, a cap set too tight can cost you money for nothing. If an OpenAI reasoning model hits max_output_tokens mid-thought, the response comes back with status incomplete, possibly before any visible text, and you still pay for the input and the reasoning. Alert on that status rather than silently retrying.
Second, on Claude, changing effort invalidates the prompt cache. Anthropic’s own example shows a switch to low producing a full cache write of 3,546 tokens instead of a cheap read. Flip effort per request and you can lose the 90% discount from our prompt caching instalment, which is a far bigger lever. Set effort once per queue and leave it alone.
Third, effort is a suggestion. Anthropic describes it as soft guidance and max_tokens as the only hard ceiling. Gemini’s docs say minimal does not guarantee thinking is off. Low effort also hurts precisely where reasoning earns its keep: multi-step coding agents, maths, long tool chains. An agent that fails at low and retries three times costs more than one run at high. Benchmarks are averages, so test on your own workload.
A one-week plan
- Log thinking tokens per request for seven days, split by task type.
- Sort tasks into three buckets: mechanical (extract, classify, format), writing and summarising, and hard reasoning (code, maths, agents).
- Run 50 to 100 real prompts from each bucket at two adjacent levels and grade the outputs blind.
- Assign one effort level per bucket and route each bucket to its own queue, which keeps caches warm.
- Set hard output caps and an alert on incomplete responses.
- Re-test whenever you change model, because defaults differ between models from the same lab.
Bull / Bear / Neutral (Not investment advice)
Bull: reasoning is a revenue multiplier for the labs. Sonnet 5.5 costs about 50% more per task than Sonnet 5 at the same list price, so token-hungry models lift revenue without a visible price rise. Bear: that revenue depends on customers never touching the dial, and benchmarkers now chart token efficiency next to intelligence. GPT-6 Astra scores within three points of Sonnet 5.5 on roughly a seventh of the output tokens, which turns efficiency into a selling point and puts that cushion under pressure. Neutral: watch default effort levels. Anthropic setting Opus 5.5 to medium by default suggests labs expect pushback on bills. Do your own research before acting on any of this.
What this means
- Thinking tokens bill at output rates, five times input, and you never see most of them.
- On GPT-6.1 Sol, medium captures most of the quality gain from low to high; max rarely pays outside hard agentic work.
- Compare models on cost per finished task, never on price per token.
- Pin effort per queue so you keep your prompt cache discount.
Pull this lever alongside routing and caching and a mid-sized workload can plausibly shed a third or more of its spend for a week of testing.
Did you know: a Claude request with thinking display set to “omitted” returns an empty thinking field, yet bills exactly as many output tokens as one that shows you a summary.
Sources
- OpenAI: Reasoning models guide
- Anthropic: Steering thinking (effort, billing, caching)
- Anthropic: Extended thinking (legacy budget_tokens)
- Anthropic: API pricing
- Google: Gemini 3 developer guide (thinking_level)
- Artificial Analysis: Claude Sonnet 5.5 benchmark (28 Sept 2026)
- Artificial Analysis: Benchmarking GPT-6 Astra (9 Sept 2026)
- DataNorth: AI effort levels explained (GPT-6.1 Sol per-effort figures)
- eesel: GPT-6.1 Sol pricing
Related on Top Tool Stack: AI on a Budget: cut your AI costs · Prompt caching