Kimi K3.1 Pro: API, Highspeed Pricing and Open Weights
Kimi K3.1 Pro is live in the API and coding tools. Compare Highspeed costs, plan limits and the November 4 open-weight target.
Moonshot AI released Kimi K3.1 Pro on October 23, 2026, alongside a Highspeed variant. The practical headline is availability: developers can use the hosted models now, while the company says full downloadable weights are scheduled for November 4. For teams evaluating a coding or research agent, the important questions are which access route works today, what speed costs, and which account restrictions matter.
What launched, and what has not launched
Kimi describes K3.1 Pro as a post-trained successor to its K3 family, with a context window of up to one million tokens and improvements targeted at coding, 3D modeling, financial research and long-running agent tasks. Both K3.1 Pro and Highspeed are available through Kimi's website and apps, Kimi Work, Kimi Code and the Kimi API. Kimi Enterprise administrators must enable the model for their organizations; the company says it is off by default at launch.
Hosted access is live, but the promised November 4 full-weight release is still in the future as of October 25. Do not treat 'open-weight model' in the launch announcement as evidence that K3.1 Pro weights can already be downloaded or self-hosted.
Regular versus Highspeed: speed and cost
Kimi says Highspeed produces output roughly two to three times faster while retaining the same model capabilities. Its current Chinese API platform lists regular K3.1 Pro at ¥13 per million input tokens and ¥68 per million output tokens. Highspeed is listed at ¥26 input and ¥136 output, with both variants showing ¥1.30 per million cached-input hits. These are yuan-denominated figures shown on the Chinese platform, not a claim about worldwide billing or every subscription plan.
Which variant fits your workload?
| K3.1 Pro | K3.1 Pro Highspeed |
|---|---|
| Lower standard API token rates in the published Chinese price list | About 2× standard input and output token rates in that list |
| Prefer for batch work or agent steps where throughput is not urgent | Consider for interactive coding where faster model output is valuable |
| Up to 1M context where the account permits it | Same stated context capability, subject to access rules |
| Measure completion quality and total cost on your tasks | Measure end-to-end latency: tools and scripts do not speed up automatically |
How developers can access it
- For pay-as-you-go API use, select kimi-k3.1-pro or kimi-k3.1-pro-highspeed in the Kimi API Platform; confirm the region's current pricing and access requirements.
- For Kimi Code, start a fresh session and select k3.1-pro or k3.1-pro-highspeed using the official model picker or /model command. Coding-plan model IDs differ from pay-as-you-go API IDs.
- Keep thinking enabled. The Kimi Code documentation says K3.1 Pro requests with thinking disabled are rejected, and supports low, high or max effort.
- Check your account tier before using very long contexts: Kimi Code lists full 1M context for Pro and higher, rather than every eligible Plus account.
- Benchmark representative prompts with identical tools and evaluation criteria before routing production work to either variant.
What the benchmark claims actually establish
Kimi reports 65.7% for K3.1 Pro on Terminal-Bench 4.0 and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, compared with 55.95% for K3 on the latter. These are company-reported measurements with specified harnesses and settings, not proof that K3.1 Pro outperforms every competing model in a particular production environment. The company also reports improved results in 3D and financial workflows; independent task-level validation remains essential.
The useful decision is not 'which model won a benchmark?' but 'what does one completed, reviewed task cost?' Run the same coding issue or document-research workflow across regular and Highspeed. Track output tokens, cache hits, time spent in model generation versus tools, acceptance rate and human correction time. If tools dominate latency, paying for faster tokens may not improve throughput enough to justify the premium.
Operational limits to check before rollout
- Model access, context size and Highspeed eligibility depend on the specific product and subscription tier.
- Kimi Code recommends new sessions when changing from older model families; switching within an existing session can affect cache behavior and quality.
- Kimi's configuration guide warns that third-party protocol conversion can change behavior; native supported API formats are the safer baseline.
- The planned open-weight release date does not establish future hosting requirements, license conditions or release-day availability.
- The company's cybersecurity risk-reduction and benchmark statements are vendor disclosures, not an independent security certification.
Questions people ask
Can I download and self-host Kimi K3.1 Pro weights today?
No. As of October 25, 2026, Kimi offers K3.1 Pro through its hosted products and API, but says the full model weights will be released by November 4. That is a stated future release target, not proof that downloadable weights or a particular license are available now. Check Kimi's official release page again before planning a self-hosted deployment.
What is the difference between K3.1 Pro and K3.1 Pro Highspeed?
Kimi describes Highspeed as the same underlying model capability with approximately two to three times faster generation. On its Chinese pay-as-you-go API price list, Highspeed's standard input and output rates are double the regular model's, while cached-input hit rates are listed equally. Faster token generation does not necessarily make tool-heavy end-to-end agent runs two to three times faster.
Does every Kimi Code account get the full one-million-token context window?
No. Kimi Code's official model table lists K3.1 Pro for Plus and higher plans, but reserves its full 1,048,576-token context entitlement for Pro and higher. Highspeed is listed for Pro and higher. Legacy plan names have different thresholds. API access and membership quotas are separate considerations, so check the exact product and account before designing long-context workflows.
Why does my K3.1 Pro coding request return HTTP 400 or 403?
A 400 response can occur when thinking is disabled or a request exceeds the model's allowed context; K3.1 Pro supports thinking effort levels but not thinking-off mode. A 403 can indicate that the account lacks access to the model or requested one-million-token context. Kimi's official configuration guide recommends checking the model ID, plan, thinking setting and context length before retrying.
Kimi K3.1 Pro is available through hosted channels now; the full weights are promised by November 4. Choose regular versus Highspeed using measured workflow economics, confirm product-specific model IDs and account limits, and avoid treating launch benchmarks or future self-hosting plans as guarantees.
Explore related Make Better resources
- AI Workflow Automation Templates— Relevant to measured agent and automation workflows
- What Should You Automate First?— Framework for choosing repeatable AI tasks
Sources & useful resources
- Kimi: Introducing K3.1 Pro— Official release; hosted availability, future full weights, benchmark methods, security claims
- Kimi Code: Model Configuration— Official plan access, model IDs, thinking mode, context limits and errors
- Kimi Code: What's New— Official October 23, 2026 launch and Kimi Code availability
- Kimi API Platform— Official Chinese-platform CNY token rates for both variants; regional pricing caveat