AI models · data verified November 9, 2026

Best AI for coding in 2026

Claude models lead the coding tests that vendors have published right now. Anthropic reports 70.6% on Terminal-Bench 4.0 for Sonnet 5.5, which costs $2 / $10 per million tokens. For less money, look at GPT-6.1 Sol or DeepSeek V4.1 Flash. If your code must stay in the EU, pick Mistral Large 4.

  1. Claude Sonnet 5.5 Anthropic · balanced

    Anthropic reports 70.6% on Terminal-Bench 4.0, a test of coding work in the terminal, the highest score in our catalogue. It costs $2 / $10 per million tokens, half the price of Opus 5.5. Watch the setting, though: Claude Code defaults to medium effort, where it scores only 31 on the Artificial Analysis index instead of 55 at max.

    API price: $2 / $10 per 1M tokens (input / output) · Context window: 1M tokens

  2. Claude Opus 5.5 Anthropic · flagship

    Anthropic reports 66.4% on Terminal-Bench 4.0, and it scores 59 on the Artificial Analysis Intelligence Index v5.0.1. Anthropic recommends it for complex, open-ended work, where it says Opus is clearly stronger than Sonnet. At $4 / $20 it costs twice as much, yet the cost per task in Artificial Analysis testing came out almost the same: $10.04 versus $10.02.

    API price: $4 / $20 per 1M tokens (input / output) · Context window: 1M tokens

  3. GPT-6 Astra OpenAI · flagship

    OpenAI's strongest model: OpenAI reports 57.9% on Terminal-Bench 4.0, and it scores 57 on the Artificial Analysis index v5.0.1. It is also the most expensive model here at $10 / $50 per million tokens. It is strict with security work and refuses things like proof-of-concept exploit code. ChatGPT Plus gives you only limited use in Codex, Pro gives you the full allowance.

    API price: $10 / $50 per 1M tokens (input / output) · Context window: 1.1M tokens

  4. GPT-6.1 Sol OpenAI · balanced

    OpenAI says it matches GPT-6 Astra on the DeepSWE v1.1 coding test at roughly a fifth of the cost. It scores 55 on the Artificial Analysis index v5.0.1, and one test task cost $1.27 against $6.24 for Astra. The API costs $2 / $10. In ChatGPT you get it from Plus upwards in Codex and ChatGPT Work, not in regular chat.

    API price: $2 / $10 per 1M tokens (input / output) · Context window: 1.1M tokens

  5. Muse Wonder Meta · flagship

    One of the few models whose Terminal-Bench 4.0 score was measured independently by Artificial Analysis: 52%, and 57% for the stronger Wonder Full. It scores 63 on the Coding Agents index in Muse Code and 44 on the main Artificial Analysis index v5.0.1. The API costs $4 / $20, and Muse Code starts at $5 a month. It is still a preview, and Meta mentions no EU data processing option.

    API price: $4 / $20 per 1M tokens (input / output) · Context window: 1M tokens

  6. Kimi K3.1 Pro Moonshot AI · flagship

    Moonshot reports 65.7% on Terminal-Bench 4.0, which would put it right behind Claude. That figure comes only from the vendor, though, so it ranks lower. On the independent Artificial Analysis index v5.0.1 it scores 46, two more than Muse Wonder. It costs $2 / $10 and has had open weights since November 2026, but running it yourself takes a data centre. API data may be stored in Singapore, and Moonshot offers no EU processing.

    API price: $2 / $10 per 1M tokens (input / output) · Context window: 1M tokens

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

    DeepSeek reports 77.0 on DeepSWE v1.1, the highest score on that test among the models in our catalogue. On Terminal-Bench 4.0, run in the Claude Code tool, DeepSeek reports only 42.3, though. It costs $0.30 / $1.20 at peak and half that off-peak, and the API plugs into Claude Code and Codex. App and API data is stored in China, so for company code, run the open weights on your own server instead.

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

  8. Mistral Large 4 Mistral AI · flagship

    For companies that want to keep both code and data in the EU. Mistral reports 41% on Terminal-Bench 4.0, a little below DeepSeek V4.1 Flash, and it scores 26 on the independent Artificial Analysis index v5.0.1. So it trails the best models, and on the tough FrontierCode 2.0 it scores 19.7% according to Mistral. But since 29 October 2026 the API costs just $0.70 / $2.10, data is processed in the EU and the weights are open. In the Vibe Pro subscription you still code with the older Medium 3.5, which Mistral will switch off in its API on 30 April 2027.

    API price: $0.7 / $2.1 per 1M tokens (input / output) · Context window: 1M tokens

Which subscription to code with

Most people do not code through the API but through a tool that comes with a subscription. Plans differ mainly in limits, not in models.

  • Claude Pro at $20 a month includes Claude Code with Opus 5.5 and Sonnet 5.5. Max from $100 only pays off if you hit the Pro limits every day.
  • ChatGPT Plus at $20 a month includes Codex with GPT-6.1 Sol and limited GPT-6 Astra.
  • Mistral Vibe Pro at $14.99 or €14.99 before VAT has Vibe CLI for all-day coding with the older Medium 3.5 model and keeps data in the EU.
  • Cheaper options: Meta's Muse Code from $5 a month and Z.ai's GLM Coding Plan from $18, which brings GLM-5.3 into Claude Code, Cline or Kilo Code.

What to watch out for in benchmarks

Most figures are published by the vendors themselves at their best settings. Different versions of a test cannot be compared: on the older Terminal-Bench 2.1, models such as Gemini 3.8 Flash score 90.8%, while on the new 4.0 even the best model is around 70%.

Price per token is not everything. Sonnet and Opus write a lot, so one Artificial Analysis task cost about $10 with them and $1.27 with GPT-6.1 Sol. The strongest Claude Fable 5.2 leads Anthropic's coding tests (49.1% on FrontierCode 2.0 against 45.5% for Opus 5.5), but Anthropic has not published a Terminal-Bench 4.0 score and it costs $8 / $40, twice as much as Opus. It pays off for the hardest long tasks, while Opus or Sonnet is enough for everyday coding. The cheap Haiku 5.5, at 39.2%, is fine for simple edits.

How we chose

The main yardstick is Terminal-Bench 4.0, a test of coding work in the terminal, because most vendors have published it in the same version. Where it is missing, we look at DeepSWE v1.1, the Artificial Analysis index and the price per million tokens and per task. Results measured independently by Artificial Analysis count for more than vendor figures, and models with unverified numbers sit lower. Where your data goes matters too.

FAQ

Which AI is best for coding?

On Terminal-Bench 4.0, Claude Sonnet 5.5 (70.6%) and Claude Opus 5.5 (66.4%). If cost per task matters, GPT-6.1 Sol came out about eight times cheaper in Artificial Analysis testing. For data kept in the EU, Mistral Large 4 is the best pick, though it trails the top models.

Is Claude better than ChatGPT for coding?

On Terminal-Bench 4.0, yes: Anthropic reports 70.6% for Sonnet 5.5, OpenAI reports 57.9% for its strongest GPT-6 Astra. Both figures come from the vendors. ChatGPT wins on price: GPT-6.1 Sol scores 55 on the Artificial Analysis index, the same as Sonnet 5.5 at max effort, and a task with Sol costs a fraction as much.

Can I code with AI for free?

Partly. ChatGPT Free includes limited Codex in the desktop app with GPT-6 Luna, which OpenAI says scores 66.6% on DeepSWE v1.1. Grok Free has limited Grok Build. Claude Free does not include Claude Code. With no limits at all, you can code with an open model on your own computer. Muse Glimmer 30B, for example, scores 76.0% on SWE-Bench Verified according to Meta.

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.