Cheapest AI API in 2026: Price per Million Tokens Compared

Updated: November 2026 · Prices checked: November 2026

API prices fell hard in 2026, and the gap between the cheapest and the most expensive model is now about 100 times. For most small-business tasks, a $0.10 model does the job, and paying more only makes sense for hard reasoning or long agent work. Below are the current prices from each vendor’s own pricing page, converted to euros, and what a typical task really costs.

How API pricing works in one minute

You pay per token. A token is a piece of a word. In English, 1,000 tokens is roughly 750 words. Bulgarian and other non-English text usually needs more tokens for the same meaning.

There are two prices. Input is what you send: instructions, the customer’s message, a document. Output is what the model writes back. Output costs 2 to 6 times more than input. Reasoning models also bill their thinking as output, so a short answer can cost more than it looks.

All vendors below list prices in US dollars per million tokens. We convert at about €0.86 per dollar. Prices are before VAT.

The price table: 18 models side by side

Standard prices for short prompts, sorted from cheapest to most expensive. Euro figures are rounded.

A few notes on the table:

• Google’s prices move soon. Gemini 4 Argon is $2/$10 through November 30, 2026, then $4/$20. Gemini 3.8 Flash is $0.75/$3.75 through December 31, 2026, then $1.50/$7.50. Google says Gemini 4 Carbon is open to paying developers at the same introductory price as Argon. We covered Carbon’s launch in Gemini 4 Carbon tops the index.
• DeepSeek charges by time of day. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays. The rest of the time, including weekends, is half price.
• GPT-6.1 Terra halved the old price. See GPT-6.1 Terra at $1 per million tokens for what changed.
• Top models are still expensive. GPT-6 Astra costs $10 input and $50 output. Use it only where cheaper models fail.

Cost per typical task in euros

Per-million prices are hard to feel. Here is what 1,000 real tasks cost. We used four common jobs:

• Sort an email: 500 tokens in, 20 out.
• Answer a customer with your FAQ in the prompt: 3,000 in, 400 out.
• Write a product description: 1,000 in, 600 out.
• Summarise a 20-page document: 15,000 in, 800 out.

The maths: one customer answer on GPT-6.1 Terra is 3,000 input tokens × $1 per million = $0.003, plus 400 output tokens × $5 per million = $0.002. Total $0.005 per answer, so $5 per 1,000 answers, about €4.30. A shop that answers 300 questions a day (about 9,000 a month) pays about €39 a month on Terra, about €3.90 on Luna or Haiku 5.5, and about €155 on Opus 5.5.

For most shops, even the expensive column is small money next to one hour of staff time. The real question is quality: does the cheap model answer well enough? That is what you test, not the price.

Hidden costs that change the ranking

The list price is only the start. Five things move the real bill:

1. Long prompts cost more. OpenAI charges 2x input and 1.5x output for the whole request above 272,000 input tokens. Claude Haiku 5.5 goes from $0.10 to $0.50 input above 100,000 tokens. Muse Ember goes from $0.10 to $0.30 input above 200,000.
2. Caching cuts repeated input. GPT-6 Luna’s cached input is $0.01 per million, GPT-6.1 Terra’s is $0.05. On Claude Opus 5.5 and Sonnet 5.5 a cache hit is 5% of the input price. DeepSeek caches automatically, and cache hits cost a fraction of a cent per million. If your prompts start the same way every time, this matters more than the list price.
3. Batch is half price. Anthropic, OpenAI and Mistral all offer batch at 50% off; Gemini’s batch tier is about half too. If results can wait hours, use it.
4. Tokenizers differ. Anthropic says its tokenizer for Claude 4.7 and later produces about 30% more tokens for the same text. Compare cost per finished task on your own data.
5. Reasoning burns output. A model that “thinks” before answering bills those thinking tokens as output. Set a low reasoning effort for simple jobs.

Also watch the extras that are not tokens. Data residency, for example: OpenAI and Anthropic both add 1.1x for regional or US-only processing. xAI adds 1.1x on its US regional endpoint.

Which API to pick for which job

Some context on the cheaper options. Meta Muse Ember is built for agents that repeat the same work all day. Mistral Large 4 has open weights, so you can also host it yourself. For the top end, Claude Fable 5.2 is the most expensive model in the table, and Anthropic itself suggests cheaper modes for many tasks.

Data location matters for some businesses. DeepSeek is a Chinese provider. Before you send customer data to any API, read where it is processed and whether it is used for training. If you use xAI, know that the Grok API now watermarks text output by default.

How to test before you switch

1. Collect 50 real examples of one task, with the answer you consider good.
2. Run them through your current model and two cheaper candidates. Most vendors offer a web playground for quick tests.
3. Score each answer: good, acceptable, wrong. Count the wrong ones.
4. Calculate cost per 1,000 tasks from the token counts the API reports, not from guesses.
5. Pick the cheapest model with an error rate you can live with. Keep the expensive one as a fallback for the cases it fails.
6. Repeat every six months. Prices in this table changed several times in 2026 alone.

Write your code so the model name is one setting, not something spread through the codebase. Then switching to a cheaper model is a five-minute change, and you keep the power to walk away when a vendor raises prices.

Read also: GPT-6.1 Terra: $1 per Million Tokens, Half the Old Price · Replit Dumps Anthropic for OpenAI Over Price – Don’t Get Locked In · AI Video Generators Compared: Price per Second and Length (2026)

Sources: OpenAI API pricing · Claude API pricing · Gemini API pricing · DeepSeek pricing · Mistral pricing · xAI pricing · Meta Muse Ember

II

About the author

Ivan Ivanov

Founder ofHawkValue

If it doesn’t save you time or money, you won’t see it here

• Facts, not hypeEvery price and number is checked against the official source
• Always the real priceYou see what it costs before you click
• I say when it’s weakEven when it’s OpenAI or Google

All articles by Ivan →