New on PrimaLabs
Moonshot AI's most capable open-weight model, with a one-million-token context. Point Claude Code, Cursor or any OpenAI SDK at PrimaLabs and make your first call in minutes.
$1.80 input · $0.18 cached input · $9.00 output per 1M tokens · 10% below Moonshot's own API list price · pay as you go
Independent check
The Artificial Analysis Intelligence Index is an independent composite of nine evaluations across knowledge work, coding and science. On it, Kimi K3.1 Pro, at max effort, scores level with GPT-6.1 Sol at medium reasoning and within two points of Claude Opus 5.5 at medium.
| Model | Reasoning setting | Index | Score |
|---|---|---|---|
| Claude Opus 5.5 | medium | 48 | |
| Claude Fable 5.1 | high | 47 | |
| Kimi K3.1 Pro | max | 46 | |
| GPT-6.1 Sol | medium | 46 | |
| GPT-6.1 Terra | xhigh | 45 | |
| Claude Fable 5.1 | medium | 43 | |
| Claude Sonnet 5.5 | high | 42 |
Top closed models at their highest settings score higher. Kimi K3.1 Pro is for the work you would otherwise send to a closed model at medium or high reasoning, on weights you can inspect.
Source: Artificial Analysis, Intelligence Index v5.0.1, accessed 5 Nov 2026. Selected models. Reasoning settings as Artificial Analysis records them. Scale on the bars is 0 to 60.
Where it fits
Kimi K3.1 Pro reasons before it answers, at low, medium, high or max effort, and you can turn thinking off. Every model below runs on the same PrimaLabs API key.
Use Kimi K3.1 Pro for
Same key, a smaller model for
Pricing
A coding agent resends its whole conversation every turn. History served from the cache bills at $0.18 per 1M, so a stable prompt prefix keeps long sessions affordable. Check prompt_tokens_details.cached_tokens in each response to see what the cache served.
Production volume
Self-hosting a 2.8T-parameter model means sourcing and running the GPUs. If you need a larger guaranteed share than serverless gives, tell us your workload and an engineer will scope dedicated capacity, throughput, latency and cost with you.
Talk to an engineerPartner in the DOE Genesis Mission with NVIDIA and national labs
FAQ
$1.80 per 1M input tokens, $0.18 per 1M cached input tokens and $9.00 per 1M output tokens, 10% below Moonshot's own API list price. Billed per token, pay as you go, with no upfront commitment.
Create an account, add balance in the dashboard and create an API key. Then point your client at PrimaLabs and set the model to primalabs-ai/Kimi-K3.1-Pro. You can make your first call in a few minutes.
Yes. Claude Code and the Anthropic SDK take https://api.primalabs.ai as their base URL (ANTHROPIC_BASE_URL). Cursor, Cline, Continue, Roo, Kilo and the OpenAI SDKs take https://api.primalabs.ai/v1. Then set the model and your PrimaLabs API key.
At max effort, Kimi K3.1 Pro scores 46 on the Artificial Analysis Intelligence Index v5.0.1, an independent composite of nine evaluations. That is level with GPT-6.1 Sol at medium reasoning and within two points of Claude Opus 5.5 at medium. Top closed models at their highest settings score higher.
Kimi K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute. Moonshot AI reports 61.2% on Kimi Code Bench 3.0 Long-horizon, its internal coding benchmark, up from 55.95% for Kimi K3. On PrimaLabs, you call it as primalabs-ai/Kimi-K3.1-Pro.
Yes. Moonshot AI publishes the weights on Hugging Face under its Kimi K3.1 Pro license. The model has 2.8T parameters, so PrimaLabs runs the GPUs and you call it per token.
Every request passes three limits, listed in the rate limits docs. If you need a larger guaranteed share than serverless gives, talk to us about dedicated capacity.
Yes. There is no subscription or upfront commitment: add balance in the dashboard and pay per token for what you use.
Caching is automatic and best-effort. A request that starts with a prefix your account sent recently is served from the cache while there is room for it. Those tokens bill at the cached rate of $0.18 per 1M. The response's usage field reports them in prompt_tokens_details.cached_tokens. Your cache is your own: no other account's request can reuse it.
Yes. On OpenRouter, request moonshotai/kimi-k3.1-pro and set provider.order to ["primalabs"] to send your requests to PrimaLabs first.
No. The serverless API runs with zero data retention for request and response content. Prompts and completions are processed in memory to serve the request and are not written to logs, databases or exports. PrimaLabs keeps request metadata only: model ID, timestamp, latency, token counts, API key and account identifiers, and HTTP status. See data retention.
Kimi K3.1 Pro has a context window of 1,048,576 tokens. On PrimaLabs you can use all 1,048,576 tokens, shared between your prompt and the response.
If you need a larger guaranteed share than serverless gives, talk to us about dedicated capacity. Talk to an engineer to scope throughput, latency and cost for your workload.
Three settings in the agent you already use. Pay per token, no commitment.
