Kimi K3.1 Pro

Strengths

• In my API sample, the model correctly returned JSON, a VAT invoice, and a memory estimation. The generated code passed 7/7 test cases.
• In a separate set of 10 simple tasks, Pro and Highspeed scored 10/10. The set was too easy to determine an advantage over K3.
• API: 2 USD per million input tokens and 10 USD output; 1M token window according to documentation.

Limitations

• With an 1800-token limit, two out of six attempts did not return a final answer. Retries with a 6000 limit were also cut off. This is a test of limits and settings, not proof of general language incompetence.
• Open weights were announced for release by November 4, 2026. I have not yet confirmed their availability.
• I did not test the app interface, games, the full 1M window, or tasks in every language here.

Use cases

I would choose the API for extraction and code work, with response validation and finish_reason checking. The small sample does not justify automatically choosing the model for every task.

Who it is for

For a developer who can adjust the token limit and measure the cost of a successfully completed task. It convinces me less with very short output limits. The tested prompts were in Polish; I do not present them as a test of other languages.

API pricing

Pro: 2/10 USD, Highspeed: 4/20 USD per million input/output tokens, cache input 0.20 USD. This is API pricing, not a subscription price. Cost of 6 Pro attempts: 0.055802 USD; Opus 5.5: 0.032004 USD; GPT-6 Astra: 0.030670 USD.

Alternatives and method

Opus 5.5 and GPT-6 Astra completed all six of the same attempts within the 1800-token limit. In a separate test with a 16000 limit, Pro scored 10/10 in 76.0 s, and Highspeed 10/10 in 32.6 s. Different prompts and settings preclude combining these results into a single ranking.

Verdict

7/10: I recommend testing K3.1 Pro for code and data, but I do not consider a low token rate a guarantee of the cheapest task. My method consisted of six single-shot calls via OpenRouter, temperature 0.2, requested reasoning low, max_tokens 1800, without streaming. A separate 10-task test used a 16000 limit and default reasoning. The lack of statistical repetitions and different conditions limit the conclusions.