Moonshot AI has introduced Kimi K3.1 Pro, the first model in its new K3.1 series, with a 1M-token context window. The company highlights improvements in coding, 3D creation, financial research and agent tasks that require sustained, multi-step work. See its official model guide.
The coding results caught my attention. In Moonshot’s reported comparisons, K3.1 Pro approaches leading proprietary models such as Claude Opus 5.5 and GPT-6 Astra. On Terminal-Bench 4.0, it sits just behind Opus 5.5 among the models shown:
| Model | Terminal-Bench 4.0 score reported by Moonshot |
|---|---|
| Claude Opus 5.5 | 66.4% |
| Kimi K3.1 Pro | 65.7% |
| GPT-6 Astra | 58.2% |
| DeepSeek V4.1 Flash 1017 | 42.3% |
These are vendor-reported results, not an independent test performed for this article. Moonshot’s technical post says it evaluated Kimi with the Claude Code agent harness for Terminal-Bench 4.0, while other scores were drawn from release reports or the benchmark leaderboard. The comparison does not establish a universal ranking for every coding task.
My takeaway is that K3.1 Pro looks like a front-runner among Chinese-developed models in this comparison. It also comes out well ahead of DeepSeek V4.1 Flash 1017 on the coding tests shown. The appeal is a potentially cheaper alternative to flagship models from OpenAI, Anthropic and Google for complex programming work; the actual cost of a finished task still depends on usage and retries.
3D modeling and financial work
The improvements extend beyond coding. According to the official guide, K3.1 Pro can create and refine complex Blender models without an MCP connector by writing and running modeling scripts. It still needs an environment capable of executing those scripts; “no MCP” does not mean “no tools.”
Moonshot also highlights stronger financial research and modeling workflows. Those are reported capabilities, rather than results I have independently reproduced here.
API pricing
The original article quotes CNY 13 per million input tokens and CNY 68 per million output tokens. These match the standard K3.1 Pro rates on the Chinese API platform. For international readers, the international API platform lists US$2 input and US$10 output per million tokens; these are separate published tariffs, not a currency conversion made here.
Caching has its own rates, so the input/output headline is not the entire billing picture:
| Standard K3.1 Pro billing item | Chinese platform, CNY per 1M tokens | International platform, USD per 1M tokens |
|---|---|---|
| Input | 13.00 | 2.00 |
| Output | 68.00 | 10.00 |
| Cache hit | 1.30 | 0.20 |
| Cache write, 5-minute TTL | 13.00 | 2.00 |
| Cache write, 1-hour TTL | 26.00 | 4.00 |
Prices were checked for this translation on October 23, 2026. See the Chinese billing details and international billing details for current terms. These are pay-as-you-go API rates for K3.1 Pro, not the separately priced Highspeed variant or a Kimi subscription.
Translated from the original Chinese article, published October 23, 2026. Original chart retained; benchmark attribution and API billing details clarified against official sources.
Comments NOTHING