Every model launch comes with a price per million tokens, and almost every comparison stops there. For a reasoning model, that's the wrong number.
AI model cost per task is what a model charges to finish one piece of work: the input, the answer, and the thinking tokens it generates along the way, which you don't see but do pay for. In October 2026, two models with exactly the same token price differ by a factor of seven on this measure. Buying on token price means buying blind.
Why does the price per million tokens tell you so little?
The price per million tokens tells you little because reasoning models do much of their work in tokens that never appear in the answer. Every major provider bills those thinking tokens as output, the most expensive rate.
OpenAI's guide to reasoning models says reasoning tokens "are not visible via the API" but still take up context and "are billed as output tokens". Anthropic is blunter in its documentation on thinking and cost: "You are billed for the full thinking process, not the thinking content visible in the response." You may see a summary of the reasoning; you pay for all of it. Google's Gemini pricing page labels the column "Output price (including thinking tokens)".
There's a second distortion. A token isn't the same size everywhere. Anthropic's model overview notes that Claude models from version 4.7 onward use a new tokenizer that produces roughly 30% more tokens for the same text. The same Dutch quote costs more tokens with one provider than another before the model has thought about anything.
What you want to know is what a model costs to finish your work correctly. No pricing page shows that number.
What do the major AI models cost per task in October 2026?
The best public approximation of cost per task comes from Artificial Analysis, an independent benchmarking firm that runs every model through the same ten evaluations (the Intelligence Index, version 4.3) and records how many tokens each one burns. The table puts the token price next to that measured cost per task, retrieved on 17 October 2026.
| Model (effort) | Price in / out per 1M tokens | Index score | Output tokens for the index | Cost per task |
|---|---|---|---|---|
| Claude Opus 5.5 (max) | $4 / $20 | 58 | 260M | $5.98 |
| Claude Sonnet 5.5 (max) | $2 / $10 | 56 | 420M | $5.46 |
| Claude Opus 5.5 (high) | $4 / $20 | 54 | 53M | $1.82 |
| GPT-6 Astra (max) | $10 / $50 | 53 | 60M | $3.26 |
| Gemini 4 Argon (high) | $2 / $10 | 53 | 110M | $1.99 |
| GPT-6.1 Sol (max) | $2 / $10 | 52 | 67M | $0.72 |
| Claude Sonnet 5.5 (high) | $2 / $10 | 47 | 52M | $0.88 |
| Gemini 3.8 Flash (high) | $0.75 / $3.75 | 41 | 170M | $1.24 |
| DeepSeek V4.1 Flash (max) | $0.30 / $1.20 | 39 | 250M | $0.27 |
| Mistral Large 4 preview | $1.36 / $4.18 | 38 | 200M | $1.13 |
| GPT-6 Luna (max) | $0.10 / $0.50 | 38 | 140M | $0.07 |
Source: Artificial Analysis model pages, for example Claude Sonnet 5.5 and GPT-6.1 Sol. Index tasks aren't business tasks, so read the last column as a ratio, not as your invoice.
Look at the four models priced at $2 / $10. Claude Sonnet 5.5 at maximum effort costs $5.46 per task, GPT-6.1 Sol $0.72. Same rate, 7.6 times the money, because Sonnet at that setting used 420 million output tokens and Sol used 67 million. The cheapest tier shows the same thing: Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 / $0.50, but Haiku used 440 million tokens to Luna's 140 million and costs three times as much per task.
A lower rate can end up dearer. Gemini 3.8 Flash costs less than half of GPT-6.1 Sol per token, yet $1.24 per task against $0.72, and it scores lower.
Effort level is the dial you control
Effort is a per-request setting on most current models, and it's the largest cost lever you set yourself. The same model at a lower effort is often a better buy than a cheaper model at maximum effort.
Claude Opus 5.5 shows the gap. At "max" it scores 58 and costs $5.98 per task. At "high" it scores 54 and costs $1.82. Four points less for under a third of the price. Sonnet 5.5 drops further: from 56 to 47, and from $5.46 to $0.88.
Check the defaults. Anthropic sets Opus 5.5 and Haiku 5.5 to "medium" and Sonnet 5.5 to "high" unless told otherwise, and DeepSeek turns thinking on at "high" by default. If your integration doesn't set the level explicitly, the provider decides what a task costs.
Our rule: start a new task at the lowest effort that passes the test set, and raise it only where the test set forces you to. How to build that test set is in our piece on evaluating AI benchmarks.
How do you work out cost per task for your own process?
You work out cost per task for your own process with four numbers: input tokens, visible output tokens, thinking tokens and the error rate. Most comparisons skip the last one, and it's usually the biggest amount.
An indicative example. A wholesaler has 20,000 incoming documents a month read and entered into its ERP. Each document is 3,000 tokens in and 500 tokens out. We compare a mid-tier model with a budget one.
Model A (rate $2 / $10, measured 1,500 thinking tokens per document):
- input: 3,000 × $2 / 1,000,000 = $0.006
- output including thinking: 2,000 × $10 / 1,000,000 = $0.020
- $0.026 per document, $520 a month, or €462
Model B (rate $0.10 / $0.50, measured 6,000 thinking tokens per document):
- input: 3,000 × $0.10 / 1,000,000 = $0.0003
- output including thinking: 6,500 × $0.50 / 1,000,000 = $0.00325
- $0.00355 per document, $71 a month, or €63
On tokens, model B wins by €399 a month. Now assume model A gets 2% of your test set wrong and model B 6%, and that an employee costing €55 an hour needs four minutes to spot and fix an error. Four extra percentage points is 800 documents a month. 800 × 4 minutes is 53 hours, or 800 × €3.67 = €2,933 of correction work. Model A comes out over €2,500 a month cheaper.
Exchange rate: the ECB reference rate of 16 October 2026, 1.1253 dollars per euro. The thinking tokens and error rates are assumptions. Measure them on your own documents: it's an afternoon's work with fifty real examples.
The reverse happens too. On a task a cheap model handles without errors, such as sorting email or extracting a fixed set of fields, an expensive model only charges you for thinking that adds nothing. That's when splitting work across models pays, as we describe under model routing with small language models.
Which discounts and surcharges change the sums?
Discounts and surcharges on AI APIs in 2026 are large enough to flip a comparison. Five matter for a company in the Netherlands.
| Factor | What it does | Who |
|---|---|---|
| Batch processing | 50% off input and output, answer not immediate | OpenAI, Anthropic, Google, Mistral |
| Off-peak rate | 50% off outside peak hours | DeepSeek |
| Prompt cache | Repeated input at 2 to 10% of the input rate | all five, with different rules |
| Long context | Higher rate above a threshold | OpenAI above 272K tokens, Claude Haiku 5.5 five times above 100K |
| EU processing | 10% surcharge | OpenAI data residency endpoints, Claude on Bedrock or Google Cloud regional, Gemini on Vertex outside "global" |
The last row hits our readers hardest. OpenAI's page on data controls charges "a 10% uplift" for data residency endpoints on models released after 5 March 2026. If you require processing in the EU, which with personal data is usually wise, that 10% belongs in every calculation. What EU data residency does and doesn't settle is covered in our piece on EU data residency with US AI providers.
Batch is the underrated one. Document processing, overnight reports and enriching a customer file don't need an answer within a second. For that work, halve the token cost in your calculation.
Which AI prices are about to change?
Several October 2026 AI prices are temporary, and the pricing pages say so. Calculate with the rate that will apply when your system is in production, not the launch rate.
- Gemini 4 Argon costs $2 / $10 until 30 November 2026 and $4 / $20 from 1 December, according to the Gemini API pricing page. The $1.99 per task in the table above was measured at the temporary rate.
- Gemini 3.8 Flash moves from $0.75 / $3.75 to $1.50 / $7.50 on 1 January 2027.
- Mistral Large 4 has a launch price of $0.68 / $2.09 against a list price of $1.36 / $4.18, according to Mistral. The page gives no end date. What the model means beyond price is in our piece on Mistral Large 4.
Doubling the token price doubles cost per task. Model A in the example above would cost €924 a month instead of €462 on Gemini 4 Argon from December, at the same usage.
When is cost per task the wrong yardstick?
Cost per task is the wrong yardstick when volume is so low that tokens don't matter, or when the data may not leave the company. In both cases something else decides.
At a few hundred tasks a month, the gap between $0.07 and $5.98 per task is a few hundred euros. Quality and error cost decide, and the most expensive model at maximum effort is often simply right. At tens of thousands of tasks a day it flips, and the question becomes whether your own model on fixed hardware is cheaper. We calculate where that break-even lies in AI without API costs.
With confidential data, the question isn't which API model is cheapest but whether an API model is allowed at all. Then the sum is local AI versus cloud AI, with hardware and operations instead of tokens. For the wider picture on model choice and buying, see our guide to hiring AI consulting. If you want cost per task worked out for your own process, error cost included, we do that as part of AI consulting.