AI token costs: frontier providers compared

TL;DR. If you calculate AI costs with input and output prices only, you calculate too short. Four price dimensions drive the bill today: input, output, thinking tokens and cache. Between the most expensive and the cheapest current model there is a factor of 100 on output, so the mix of models matters more than the discount you negotiate. Prompt caching is the biggest single lever: with Claude Opus 5.5 and Sonnet 5.5, a cache hit costs only 5% of the input price. And watch out: Gemini 4 Argon prices double on December 1, 2026.

Four price dimensions in 2026

Two years ago, a token calculation counted two dimensions: input and output. Today there are four, and the two new ones often decide the larger part of the bill. All prices in this post are list prices in US dollars per million tokens, as of November 3, 2026.

• Input tokens ($0.10 to $10). Everything in the prompt: system prompt, message, tool definitions, context from the knowledge base, history. With agents, input grows with every step because the whole history is sent again.
• Output tokens ($0.50 to $50). The model’s answer. For the current models from Anthropic, OpenAI and Google, output costs five times as much as input. For JSON, code and reports it is the biggest driver; for classification and short answers it barely matters.
• Thinking tokens (reasoning). Many models think before they answer. Users never see these tokens, but they are billed at the output price. OpenAI says so explicitly, and Google labels its output price as “including thinking tokens”. How many there were is shown in the usage field of the API response.
• Cache tokens. The biggest single lever. If you send the same beginning of a prompt again and again, you can cache it. At Anthropic, writing to the cache costs 1.25 times the input price (five minutes) or twice the price (one hour); reading it afterwards costs only 10% of the input price. For Opus 5.5 and Sonnet 5.5 it is just 5%.

A caching example: a 12,000-token system prompt is read 100 times an hour with Sonnet 5.5. Without a cache that is 1.2 million input tokens, or $2.40 per hour. With a one-hour cache, writing once costs $0.05 and reading 99 times costs $0.12. Together that is $0.17, roughly one fourteenth.

Provider pricing in 2026 compared

Since May, the price list has been renewed almost entirely. The spring models, such as Claude Opus 4.7, GPT-5 or Gemini 2.5, have been replaced. From December 1, 2026, Google routes requests for Gemini 2.5 Pro and Flash to newer models.

Sources are the pricing pages of Anthropic, OpenAI (prompts up to 272,000 tokens), Google and Mistral. Thinking tokens are included in the output price everywhere. Haiku 5.5 costs five times as much for prompts over 100,000 tokens. Enterprise terms and prices through cloud marketplaces may differ.

Hidden costs: what no price sheet shows

Three cost drivers appear in no price table. They are the most common source of blown budgets.

• Images and PDFs cost more than expected. Claude splits an image into patches of 28 by 28 pixels, each patch is a token. An image of 1000 by 1000 pixels therefore costs 1,296 tokens, a full HD screenshot on the newer models 2,691. A PDF page costs 1,500 to 3,000 tokens for the text plus the tokens for the page image. A 40-page PDF quickly runs into six figures, and without a cache that is charged again on every request.
• Tool calls create several waves of tokens. With a tool call, the model first writes the call (output), gets the result back (input) and then writes the answer (output). Tool definitions and results count as input, even if a result is 5,000 lines of JSON. With every further step, the history so far is sent again.
• Thinking tokens and new tokenizers. A model that thinks for a long time costs many times what the visible answer suggests. There is also a quiet effect: from Claude 4.7 onwards, a new tokenizer produces about 30% more tokens for the same text. A comparison based only on the list price per token misses that.

Tier mix strategy

Sending every task to the most expensive model is costly and rarely necessary. Most everyday requests are classifying, routing and extracting. Small models are enough for that. A sensible assignment looks like this:

For the worked example below, we assume a starting point of 60% small models, 30% mid tier and 10% top models by token volume. Which mix works for you only shows when you measure it on your own tasks. Small models make more mistakes on hard cases. Those mistakes cost money too.

Worked example: a machine builder with a tier mix

A mid-sized machine builder with around 180 employees uses AI for three tasks: questions to the knowledge base, drafts for quotes and pre-sorting service tickets. Together that is 2.4 million tokens a day. We assume 2.1 million of them are input and 0.3 million output.

• Everything on Opus 5.5: 2.1 × $4 plus 0.3 × $20 gives $14.40 a day.
• Tier mix 60/30/10 with Haiku 5.5, Sonnet 5.5 and Opus 5.5: input costs $1.06 and output $5.30 per million tokens on average. That gives $3.82 a day.

Over 22 working days that is $316.80 versus $83.95 a month. The factor matters more than the amount: it holds at ten times the volume as well. Add caching, for example for the system prompt of the ticket pre-sorting, and input costs fall further, as the example above shows. On top of that, the batch API at Anthropic and OpenAI halves the price of anything that does not need to be ready at once, such as overnight analyses.

FinOps discipline: what production AI needs at minimum in 2026

• Token logging per request. Every request records model, input, output, thinking and cache tokens, user and use case. Without that log there is no cost attribution and no lever.
• Budget alerts per day and week. Thresholds per use case and team. A warning at 80% and a hard stop at 120% of the daily budget prevent an agent from burning money in a loop.
• Budgets for users and teams. A cap per user protects against misuse and bugs. It is enforced centrally in an API gateway, not in every application separately.
• Cache rate as a metric. The share of input tokens that comes from the cache. If you do not measure it, you will not notice when a small prompt change quietly breaks the cache.
• Price changes in the calendar. The doubling of Gemini 4 Argon on December 1, 2026 and the rerouting of older Gemini models show that prices and models change faster than contracts. Check the pricing pages at least once a quarter.

Conclusion

Token costs are a question of FinOps discipline, not of procurement. If you calculate without the four price dimensions, without a tier mix and without a caching strategy, the worked example shows you pay 3.8 times as much, with comparable results for most tasks.

What you should do now:

1. Break down your own bill. What share is input, output, thinking, cache? The API’s usage data shows it for every request.
2. Check the Gemini date. If you use Gemini 4 Argon or still Gemini 2.5, recalculate your costs from December 1, 2026.
3. Turn on caching for system prompts, tool definitions and long documents that repeat.
4. Test a tier mix. Move the simple tasks to a small model and measure the error rate on real cases.
5. Set budgets before the first agent goes into production.

Prices are vendor list prices as of November 3, 2026. If you would like to go through your token usage with us, we are happy to talk.