AI models · data verified October 28, 2026

Cheapest AI API in 2026: price and quality

Claude Haiku 5.5 offers the best value: $0.10 per million input and $0.50 per million output tokens, and first place in its class on the Artificial Analysis index. For a similar price with data in the EU, use GLM-5.3-Flash through Mistral or Qwen3.8-Flash in Frankfurt.

  1. Claude Haiku 5.5 Anthropic · fast and cheap

    $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens, and 36 points on the Artificial Analysis index, first of 22 comparable models. Longer prompts cost five times more ($0.50 / $2.50), though, the new tokenizer splits the same text into about 30% more tokens than Haiku 4.5 did, and the model writes long answers. The direct API stores data in the US.

    API price: $0.1 / $0.5 per 1M tokens (input / output) · Context window: 1M tokens

  2. GLM-5.3-Flash Z.ai (Zhipu AI) · fast and cheap

    $0.15 / $0.50, and Z.ai cites 57 points on the Artificial Analysis index. That is the older v4.1.1 version of the index, though, which cannot be compared directly with today's v5.0. Through the Mistral API it costs the same and data is stored in the EU. It takes text, images and video and has a 1-million-token context. Thinking cannot be turned off, so even a simple answer is not instant.

    API price: $0.15 / $0.5 per 1M tokens (input / output) · Context window: 1M tokens

  3. GPT-6 Luna OpenAI · small

    As cheap as Haiku at $0.10 / $0.50, with the highest rate limits in the GPT-6 family. It scores only 28 on the Artificial Analysis index, though. OpenAI reports 66.6% on the DeepSWE v1.1 coding test. EU data processing costs 10% extra. Prompts over 272,000 tokens cost $0.20 / $0.75.

    API price: $0.1 / $0.5 per 1M tokens (input / output) · Context window: 1.1M tokens

  4. Qwen3.8-Flash Alibaba (Qwen) · fast and cheap

    $0.15 / $0.47, the cheapest output in this ranking, with a 1-million-token context and image and video input. In Frankfurt you get it in EU scope for the same price as in Singapore: data is stored in Germany and processing stays inside the EU. So far only Alibaba has published test results, for the open preview version Flash-Next, and nobody has reproduced them independently.

    API price: $0.15 / $0.47 per 1M tokens (input / output) · Context window: 1M tokens

  5. Mistral Small 4 Mistral AI · fast and cheap

    $0.15 / $0.60, and half that through the Batch API. A European vendor with data in the EU, an Apache 2.0 license and only 6.5 billion active parameters per answer. A 256K-token context, and it understands images. Mistral has not published a full results table, but says it beats gpt-oss-120b on LiveCodeBench.

    API price: $0.15 / $0.6 per 1M tokens (input / output) · Context window: 256K tokens

  6. DeepSeek V4.1 Flash DeepSeek · fast and cheap

    $0.30 / $1.20 at peak, half that off-peak, and cached input from $0.003. DeepSeek reports 77.0 on DeepSWE v1.1. Peak hours include 06:00–10:00 UTC on weekdays, the European morning, so daytime costs are harder to predict. API data is stored in China, and Czechia's NÚKIB rates the threat as high. Not for sensitive data.

    API price: $0.3 / $1.2 per 1M tokens (input / output) · Context window: 1M tokens

  7. GPT-6.1 Terra OpenAI · fast and cheap

    When the cheap models are not enough: $1 / $5 and 51 points on the Artificial Analysis index, 12th of 52. It costs half as much as GPT-6.1 Sol, and thinking can be switched off completely, which saves time and money on simple tasks. It is wordy, though, and you pay for output. EU processing costs 10% more in the API.

    API price: $1 / $5 per 1M tokens (input / output) · Context window: 1.1M tokens

  8. Grok 4.3 xAI · fast and cheap

    Grok 4.3 is the cheapest general-purpose Grok at $1.25 per million input tokens and $2.50 per million output tokens, so its output is cheaper than GPT-6.1 Terra's. It handles a million tokens of context. It is an older generation from April 2026, though, with no independent Artificial Analysis score, so it ranks lower. The newer Grok 4.7 costs $2 / $6.

    API price: $1.25 / $2.5 per 1M tokens (input / output) · Context window: 1M tokens

Why price per token is not everything

You pay for every token the model writes. Haiku 5.5 produced 300 million output tokens in Artificial Analysis testing, the most in its class, and GPT-6 Luna 110 million. Gemini 3.5 Flash-Lite has pricier output ($2.50) but writes briefly (44 million) and is one of the fastest models. It is weak at harder tasks, though, with 12 points on the index.

Caching saves a lot: if you keep sending the model the same opening, such as long instructions or a document, the repeated input costs a fraction of the price. With Haiku 5.5 it is $0.01 per million tokens.

Even cheaper, with a catch

  • Llama 4 Scout costs $0.08 / $0.24 at the cheapest provider. Its license grants no rights to companies headquartered in the EU, though, so an EU company may not run it or build products on it itself.
  • Ministral 3 14B costs $0.20 per million input and output tokens with data in the EU. It is an older generation, though, and Mistral lists no thinking mode for it.
  • The Gemini API has a free tier for 3.5 Flash-Lite and 3.8 Flash, and Google does not train on API data from users in the EEA.
  • Grok Build 0.1 is the cheapest Grok in the API at $1 for input and $2 for output. It is a model built for coding, though, and xAI does not offer it in an EU region.
  • Nemotron 3.5 Super VL usually costs $0.15 for input and $0.50 for output with providers on OpenRouter, and it also understands images and video. These are provider prices, though, as NVIDIA has no price list of its own, and an independent quality score is still missing.
  • Beam 1.0 costs $0.25 for input and $1 for output with the cheapest provider. Coding results so far come only from Reflection AI, and reasoning cannot be turned off, so it uses more tokens.

How we chose

We look at the price per million input and output tokens in the base tier. Among the cheap models, quality decides, measured by the Artificial Analysis index where available and otherwise by vendor tests, which count for less. We also weigh long-prompt surcharges, peak pricing, how much text a model writes and where your data goes. Models that EU companies may not run or build on under their license are left out of the ranking.

FAQ

What is the cheapest AI API?

Purely on price, Llama 4 Scout at $0.08 / $0.24, but under its license, companies headquartered in the EU may not run it or build on it themselves. Among current models of good quality, the cheapest are Claude Haiku 5.5 and GPT-6 Luna at $0.10 / $0.50 per million tokens.

Which cheap API keeps data in the EU?

GLM-5.3-Flash and Mistral Small 4 through the Mistral API, which stores data in the EU by default. Qwen3.8-Flash in Frankfurt in EU scope. GPT-6 Luna offers EU processing for 10% extra. The direct Claude API has no EU region.

Can I use an AI API for free?

For testing, yes. The Gemini API has a free tier for 3.5 Flash-Lite and 3.8 Flash, and Alibaba gives you 1 million tokens for 90 days with Qwen3.8-Max in Singapore. For production at larger volumes, you pay per token.

Sources

Prices and models change often. The data comes from official vendor pricing pages and documentation, and every page shows when it was last checked. Before a larger purchase, confirm the price with the vendor.

← All AI models

Not sure which model fits your business?

I can help you pick a model and plan based on what you actually need, set it up safely for company data and train your team to use it well. Let’s start with an intro consultation.

See pricing [email protected]