AI model comparison 2026: API pricing and which model to use in your business

There is no single “best AI model”: there are tasks, and they are almost always solved by a cheaper model than people think. This AI model comparison collects the official API prices for Claude, GPT-6, Gemini, DeepSeek and Mistral as of October 3, 2026, and says which one fits each business job.

This is a living page. We update it every time a model is released or a price changes, and at the end we note what changed and when.

The short answer

The five providers sell the same thing in three tiers, and the price difference between the bottom and the top is a hundredfold.

• Small tier (GPT-6 Luna, Gemini 3.5 Flash-Lite, DeepSeek Flash, Claude Haiku 4.5): for reading, classifying and extracting data at scale. This is where most business automation lives.
• Middle tier (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash, Mistral Large): for writing, summarizing and answering with judgment. Claude Sonnet 5.5 and GPT-6.1 Sol cost the same.
• High tier (Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1): for coding, long multi-step jobs and anything the middle tier cannot solve.

The rule we use: start with the small tier, test with real cases, and only move up if it fails. Moving back down almost never happens, so it is better not to start at the top.

API prices, model by model

List prices in dollars per million tokens, as published by each provider. Taxes not included.

The sources are each provider’s pricing pages, linked at the end. OpenAI also keeps GPT-6 Sol at the same price as GPT-6.1 Sol; in a new project, choose 6.1. Gemini 4 Argon is not in the table because you can’t use it yet.

Gemini 4 Argon: what is known

Gemini 4 exists. Google introduced it on September 30, 2026 under the name Gemini 4 Argon, as its new frontier model, but as of October 3 it cannot be used from the open API and does not appear in the Gemini model list.

• Who has it. A group of trusted cybersecurity defenders, within Google’s Fairwind program, as well as Google’s own internal teams.
• When it opens. No date. Google says it will open it as soon as possible, starting with paid API customers and Google AI Ultra subscribers, once it finishes testing its safeguards.
• How much it will cost. It will launch with a price of $2 per million input tokens and $10 for output, with cached input 95% cheaper. After that period ends, it will move to $4 and $20.
• What Google presents it for. Long multi-step tasks: programming, specialized office work such as legal or financial, and defense against cyberattacks. It can generate up to one million output tokens in a single response.

For a company automating today, nothing changes: build with the models that are already available and leave it ready to switch models by changing one setting. When Argon opens, we will add it to the table with its prices.

How to read a token price

The API is not paid per month or per user: it is paid for processed text, measured in tokens. A token is a piece of a word. As a reference, Anthropic estimates that one million tokens are about 555,000 words in its current models.

• Input is what you send: the instructions and the document.
• Output is what it answers. It costs between three and eight times more than input, depending on the model.
• Reasoning counts as output. Models that think before answering bill that reasoning even if you do not see it. Google says so in its pricing: output price includes reasoning tokens.
• Each provider counts tokens differently. The same text does not produce the same number everywhere. Anthropic warns that its recent models generate around 30% more tokens than previous ones for the same text.

That is why the price per million is useful for ranking, not for budgeting to the cent. The real cost comes from testing with your own documents.

How much it costs in practice

Two typical jobs, calculated with the prices in the table. The assumptions are ours and are shown so you can replace them with your own:

• Read 1,000 invoices and return their data: 2,000 input tokens and 300 output tokens per invoice, with the text already extracted from the PDF.
• Write 1,000 product listings: 800 input tokens and 600 output tokens per listing.

This is a calculation using list prices, not a measurement. In reasoning models, the real output is larger than the visible answer and the bill goes up. If the invoice comes in as an image and not as text, that too.

The takeaway that does hold: the same job costs cents or tens of dollars depending on the tier. Model choice matters a hundred times more than any discount.

Which model to use for each task

This is the starting point we use when setting up a process. It is not a quality ranking: it is where to begin testing.

The test that decides is always the same: twenty real cases of yours, with the correct answer already known, run through the small model. If it gets the ones it should get right, there is nothing else to buy.

Four ways to pay less

1. Batch processing. If the work is not urgent, Anthropic, OpenAI, Google and Mistral charge half price for sending it in batches. An overnight batch of invoices or records is the textbook case.
2. Instruction cache. When all requests repeat the same long instructions, that part is charged at a steep discount from the second time on. In Claude, at one tenth of the entry price or less.
3. The small model first. Let the small one solve the easy part and pass on to the medium one only what it is not clear about. This is usually the biggest discount of all.
4. Short outputs. Output is the expensive part. Asking for the data in a fixed structure and without explanations makes it cheaper and, as a bonus, easier to check.

What the price does not say

• What they do with your data. Gemini has a free tier, but Google says that in it the content is used to improve its products; in the paid tier, it is not. With customer or staff data, the free tier is not an option. Before sending personal data to any provider, you need to check its data processing agreement and where it is processed; DeepSeek is a Chinese company and Mistral is French.
• Preview is not stable. Gemini 3.1 Pro is still in preview. A model like that can change or be withdrawn with little notice, and it is not a good basis for a process that has to run on its own.
• Very long documents cost more. OpenAI doubles the entry price in long context, and Gemini 3.1 Pro doubles it from 200,000 tokens onward.
• Prices move. Gemini 3.8 Flash doubles its price on January 1, 2027, and DeepSeek charges double during its peak hours, which fall in the early morning and morning in Spain.
• The model is not the project. The provider bill is usually the small part. What costs money is connecting it to your program, handling edge cases and deciding what a person reviews.

Frequently asked questions

How much does the ChatGPT API cost?

It depends on the model. As of October 3, 2026, GPT-6 Luna costs $0.10 per million input tokens and $0.50 output; GPT-6.1 Sol, $2 and $10; GPT-6 Astra, $10 and $50. You pay per use, with no fixed fee.

How much does the Claude API cost?

Claude Haiku 4.5 costs $1 per million input tokens and $5 output; Sonnet 5.5, $2 and $10; Opus 5.5, $4 and $20; Fable 5.1, $10 and $50.

How much does the Gemini API cost?

Gemini 3.5 Flash-Lite costs $0.30 input and $2.50 output per million tokens. Gemini 3.8 Flash, $0.75 and $3.75 until December 31, 2026, and double from January. There is a free tier with limits, in which Google uses the content to improve its products.

Does Gemini 4 exist?

Yes. Google introduced Gemini 4 Argon on September 30, 2026, but as of October 3 you still can’t use it: only a group of cybersecurity defenders uses it and it does not appear in the Gemini API. It will first open to paid API customers and Google AI Ultra subscribers, with no date. The latest one that can be used today is Gemini 3.8 Flash. Gemma 4 is something else: Google’s family of open models.

How much will Gemini 4 Argon cost?

Google has announced a launch price of $2 per million input tokens and $10 output, which will rise to $4 and $20 when that period ends. It has not said how long it will last.

Is a ChatGPT or Claude subscription the same as the API?

No. The subscription is for a person to use the chat, with a fixed fee. The API is for a program to call the model, and you pay by tokens. To automate a process, you need the API.

Which AI model is best for a company?

The cheapest one that solves your task with your documents. For extracting data and classification, the small tier is usually enough; for writing, the medium one; the high one is reserved for programming and multi-step work.

Do the prices include VAT?

No. They are list prices in dollars and do not include taxes. Whether tax applies depends on the provider and where your company is located.

How often do the models change?

Several times a year for each provider. That is why it is worth making sure the process does not depend on a specific model: changing it should mean adjusting a setting, not redoing the work.

Change log

• 3-10-2026. First version, with the list prices of Anthropic, OpenAI, Google, DeepSeek and Mistral checked that day, and what Google has announced about Gemini 4 Argon, which you can’t use yet.

Conclusion

Choosing a model is choosing a tier, and the tier is decided by the task. Start with the small one, test with your cases, and only move up if it fails. The rest of the comparison changes every few months, which is why we keep it up to date here.

The prices are the list prices published by each provider on the indicated date and may change. The cost examples are indicative calculations, not a quote.

Useful links

• Automate invoices with AI: a complete process, step by step, with a real case.
• Process automation for e-commerce: what is automated, how we work, and how much it costs.
• Claude API pricing (Anthropic).
• OpenAI API pricing.
• Gemini API pricing (Google).
• Gemini 4 Argon announcement (Google, in English).
• DeepSeek API pricing.
• Mistral pricing.