Kimi K3.1 Pro: Specs, $2/$10 Pricing, Benchmarks and the Open-Weights Question

Moonshot AI shipped Kimi K3.1 Pro on October 23, 2026 (the release date Artificial Analysis lists), about three months after Kimi K3 (July 2026). It is a post-trained version of K3, not a new base model. Moonshot says it spent substantially more post-training compute on it, running joint reinforcement learning across hundreds of downstream domains. The result is a model that sits near the top of independent leaderboards, costs a third less than K3 per token, and lists for roughly half of Claude Opus 5.5's per-token price.

Two things make this release harder to evaluate than a normal model launch. First, the "open-weight" label is still a promise. Moonshot's launch post says the weights will be released by November 4, but on the date of writing nothing is downloadable. Second, a set of unverified paid-influencer allegations surfaced around the launch. Neither changes what the API does today, but both are reasons to lean on independent measurements rather than launch-week social posts.

This guide covers the specs and price table (with worked cost examples), the API quirks that break code written for K2, the benchmarks (vendor-reported and independent, labelled separately), a short comparison with Claude Opus 5.5, the open-weights question including China's draft export controls, and what K2 users need to do before the year-end retirement. If you are new to the K3 family, start with our Kimi K3 complete guide.

Kimi K3.1 Pro at a glance

K3.1 Pro is also available in the Kimi web and mobile apps, the Kimi Work desktop app (version 3.1.0 or later) and Kimi Code, where you pick it with the /model command. For Kimi Enterprise customers it is off by default, and an administrator has to enable it.

Pricing: K3.1 Pro vs Highspeed vs K3

These are the list prices on Moonshot's Kimi API pricing page as of October 2026, in USD per million tokens. Prices exclude taxes and may change.

Points to note:

• K3.1 Pro is cheaper than K3. Input, output and cached input are all about a third lower. If you are on K3 today, moving to K3.1 Pro lowers your per-token bill.
• Highspeed costs double. Input, output and cache writes are all 2x the standard model's rates. Cached input is $0.20 on both, so the gap is smaller for workloads that are mostly cache hits.
• Cache writes are a separate line item. Caching is automatic, with a default 5-minute TTL. Writing a prefix into the cache is billed at the write rate for the TTL you choose. A 5-minute write costs the same as uncached input, and a 1-hour write costs double. Each cache hit refreshes the entry's TTL. You set the TTL with prompt_cache_options.ttl (5m or 1h).
• No context-length tiers. The quickstart says pricing is flat across the full 1M window.
• Partial blocks miss the cache. Caching works in blocks. A trailing partial block is billed as a cache miss.
• Changing reasoning effort mid-session invalidates prefix-cache hits, according to Moonshot's parameter reference. Pick a level before the conversation starts and keep it.

For a deeper look at K3-family cost planning, see our Kimi K3 pricing and API costs breakdown and our explainer on how prompt caching reduces AI cost.

Worked example: an agent workload

Assume 1,000 requests per day. Each request sends 50,000 input tokens, of which 40,000 are a warm cached prefix (system prompt, tool definitions, repository context) and 10,000 are new. Each request returns 5,000 output tokens, counting reasoning. Cache-write costs are left out here for simplicity and covered below.

Output dominates in every column. With the cached prefix already warm, output is 64% of the K3.1 Pro bill ($0.050 of $0.078). That matters because K3.1 Pro is a verbose reasoner. Artificial Analysis recorded 150M output tokens on its Intelligence Index run, against a median of 87M. In practice, reasoning_effort is the strongest cost lever you have. Test low and high before defaulting to max.

Cache writes add a cost each time the prefix is written. Writing the 40,000-token prefix once costs 40,000 x $2/1M = $0.08 with the 5-minute TTL, or 40,000 x $4/1M = $0.16 with the 1-hour TTL. With steady traffic, every hit refreshes the 5-minute TTL and the write is rare. Pay for the 1-hour tier only if requests that share the prefix are regularly more than five minutes apart.

API quirks that break K2-era code

The Kimi API is OpenAI-compatible. The quickstart uses the OpenAI Python SDK with base URL https://api.moonshot.ai/v1 and an API key from platform.kimi.ai. An Anthropic Messages-compatible endpoint is also documented. Swapping only the model string is not enough, though. K3.1 Pro has the same parameter constraints as K3, and they differ from K2.6.

Thinking is always on

K3.1 Pro always reasons, with Preserved Thinking always on. There is no non-thinking mode. Moonshot's FAQ answers "How do I turn off Kimi K3.1 Pro's chain-of-thought?" with "You can't." The K2.x thinking object (including {"type": "disabled"}) is not a K3-family parameter, and the migration guide says to remove it. To reduce latency, set reasoning_effort to low.

Fixed sampling parameters

These values are fixed: temperature=1.0, top_p=0.95, n=1, presence_penalty=0 and frequency_penalty=0. Moonshot's parameter reference says "passing any other value returns an error." Many shared LLM wrappers send temperature=0 by default, and those requests will fail. Strip these fields from your requests rather than setting them.

Return the full assistant message

In multi-turn conversations and tool loops, append the complete assistant message the API returned, including reasoning_content. Do not keep only content. Streaming returns separate reasoning_content and content deltas. For structured output (response_format with json_schema and strict: true), parse only message.content.

Other constraints

• tool_choice accepts auto, none and required. required was not supported on K2.6 and K2.7 Code.
• Vision input does not accept public image URLs. Use base64 data URLs or ms://<file-id>, and send content as an array of objects.
• Moonshot says its built-in web search is being updated and is not recommended for production for now.
• Rate limits (concurrency, RPM, TPM, TPD) depend on your cumulative top-up tier.
• In Claude Code, the troubleshooting guide says to use the compatible alias kimi-k3[1m]. Direct API calls and Codex use kimi-k3.1-pro.

Minimal call

This example follows Moonshot's quickstart. Note that it sends no temperature, top_p or thinking field.

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

stream = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="low",
    messages=[{"role": "user", "content": "Explain why the sky is blue."}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta
    reasoning = getattr(delta, "reasoning_content", None)
    if reasoning:
        print(reasoning, end="", flush=True)
    if delta.content:
        print(delta.content, end="", flush=True)

Platform and plan notes

The global developer platform is platform.kimi.ai, which hosts the docs, console and API keys. Moonshot's troubleshooting guide says users outside mainland China should use platform.kimi.ai, and that accounts and keys from different regional Kimi platforms are separate and cannot be mixed. If an older key returns authentication errors, check that its region matches the endpoint.

The API is pay-as-you-go and separate from Kimi membership and Kimi Code subscriptions. A Kimi Code plan does not give you API credit, and API credit does not unlock Kimi Code plan features. Per the Kimi Code docs, K3.1 Pro is available on Moderato / Plus plans and above, while the 1M context window and K3.1 Pro Highspeed require Allegretto / Pro and above. Kimi Code's /login command signs you in to the Kimi Code platform under your membership plan, and Moonshot's troubleshooting guide says Kimi Code keys and Open Platform keys are not interchangeable. Check the current plan table before you buy seats for a team.

Benchmarks: what Moonshot reports and what independent tests show

Benchmark numbers for this release come from two places, and they should not be mixed. The tables below label each source.

Moonshot-reported results

All figures below come from the benchmark data published with Moonshot's launch post, with every model at max effort. Moonshot ran K3.1 Pro itself and took some competitor scores from official leaderboards or system cards, and the harnesses differ by benchmark, so these are not controlled head-to-head trials.

Moonshot also reports internal results: 61.2% on Kimi Code Bench 3.0 Long-horizon (up from 55.95% for K3), and gains on internal finance evaluations. On factuality, Moonshot says the claim-level hallucination rate on its internal Kimi Hallu Bench Hard (run with tools) fell from 23.4% for K3 to 8.3% for K3.1 Pro, and the share of responses with at least one factual error was 42.4% of K3's rate. VentureBeat reports the same 23.4% and 8.3% figures. Moonshot describes the benchmark as a stress test built from challenging, de-identified prompts drawn from practical work tasks and deliberately crafted to probe situations where models are prone to errors, and says "the reported rates do not represent overall error rates in daily use."

The vendor numbers tell a consistent story. K3.1 Pro roughly matches Opus 5.5 on Terminal-Bench 4.0, leads on long-horizon software and office tasks in Moonshot's own runs, and trails clearly on scientific terminal work (Terminal-Bench Science 0.1) and CAD with tools.

Independent results (Artificial Analysis)

VentureBeat also reported several of Artificial Analysis's per-benchmark results:

• Terminal-Bench 4.0: 67%, up from 13% for K3. This is close to Moonshot's own 65.7%.
• GDPval-AA: 1,846 Elo, third behind Opus 5.5 (1,866) and Wonder Full (1,857).
• K3.1 Pro ahead of Opus 5.5: AutomationBench-AA (73% vs 70%).
• K3.1 Pro behind Opus 5.5: Humanity's Last Exam (52% vs 61%) and CritPt (25% vs 32%).

Leaderboard positions change as models are added, so treat the rank as a snapshot. For how the K3 family compares more broadly, see our Kimi K3 benchmarks comparison.

A note on launch-week sentiment. VentureBeat reported that an X user posted screenshots of what he described as a brief for a paid Kimi promotional campaign. The brief allegedly told creators not to use the paid-partnership label and to frame Kimi as an open-source equivalent of Opus 5.5. The document has not been authenticated. It is unclear who wrote it or whether any paid posts were published, and VentureBeat had found no Moonshot response as of October 25. Whatever the truth, it is a good reason to base decisions on third-party measurements and your own evals, not on social media threads.

Kimi K3.1 Pro vs Claude Opus 5.5

On Artificial Analysis's measurements, Opus 5.5 is the stronger model by three index points, and it produces output nearly twice as fast. K3.1 Pro costs $2.11 per index task against $5.98 for Opus 5.5. That is about 35% of Opus's cost ($2.11 / $5.98 = 0.353), so Opus costs about 2.8x more per task ($5.98 / $2.11 = 2.83). Per-task cost is a more honest comparison than list price, because it reflects how many tokens each model actually uses.

For full Opus details, see our Claude Opus 5.5 complete guide. For the previous generation of this matchup, see Kimi K3 vs Claude Opus 5.

The open-weights question

As of October 26, 2026, the position is this:

• Moonshot says in its launch post and quickstart that the full weights will be released by November 4, 2026. VentureBeat reports that Moonshot said they will ship under an open license. Moonshot has not named the license.
• Hugging Face has no K3.1 Pro repository. The newest model under the moonshotai organization is Kimi-K3, released under the custom Kimi K3 License.
• Artificial Analysis currently lists K3.1 Pro's weights as not publicly available.

The weights are promised, not released. Until they are, treat K3.1 Pro as an API-only model for planning purposes. We will update this section once the November 4 date passes.

What Moonshot changed for the open-weight release

Moonshot says it removed cyber-related data from both mid-training and RL for the open-weight release, while keeping general-purpose coding tasks. Its in-house exploit benchmark, xdev-bench-hard (pass@3), shows the effect. A checkpoint trained without those removals scored 90.5%. The released K3.1 Pro scored 81.0%. K3 scored 33.3%. Moonshot's stated principle is not to release open weights that extend the absolute frontier of cyber capability. Security teams evaluating the model for defensive work should expect deliberately reduced offensive capability.

China's draft export controls

On October 16, 2026, The Decoder reported a draft from China's commerce and science ministries that would control 23 AI technologies. For large models, the controls apply once a model meets any one of three thresholds: 10^26 FLOPs of training compute, at least 2 trillion parameters, or a global top-10 ranking on public leaderboards. Controlled items include weights, checkpoints, source code and architecture.

K3.1 Pro has 2.8T parameters and ranks in Artificial Analysis's top five, so it appears to meet at least two of those thresholds. The draft is not a ban on open releases. It exempts technology that is already public, and it offers fast-track "automatic licensing" for publishing restricted technology as open source. It does add a registration step and gives Beijing discretion. Dean W. Ball, quoted by The Decoder, said the framework establishes "the optionality to shut down future open-weight models at will." We do not know how or whether it will affect Moonshot's November 4 date, but it is a real risk to the timeline. If your roadmap depends on self-hosting K3.1 Pro, plan around K3 (already open) until the weights are actually published.

If and when you self-host

Kimi K3's model card is the best guide to what K3.1 Pro's weights will likely look like. It lists 2.8T total and 104B activated parameters, MXFP4 weights with MXFP8 activations from quantization-aware training, and deployment recipes for vLLM and SGLang. A model this size needs multi-GPU inference infrastructure, plus engineers who can manage quantization, serving throughput and evaluation (Codersera can help you hire vetted AI developers with LLM serving experience). Read the license terms when they are published before you build a product on them. For the wider open-model landscape, see our best open-source LLMs of 2026 roundup.

K2 deprecation: what K2 users need to do

Moonshot's pricing and model-list pages carry this notice. Starting October 31, 2026, K2 series models will no longer be available to newly registered API users. Existing users are not affected for now. Moonshot plans to retire all K2 series models by the end of December 2026. The K2 models still listed are kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6, all with a 262,144-token context window.

Moving from K2.7 Code to K3.1 Pro raises input pricing about 2.1x ($0.95 to $2.00, since 2.00 / 0.95 = 2.11) and output pricing 2.5x ($4.00 to $10.00). Because output dominates agent bills, expect your total to rise by roughly 2-2.5x. Re-run your cost model with your real token counts before the cutover, not after.

Migration checklist:

1. Change model to kimi-k3.1-pro (or -highspeed).
2. Remove the thinking object, and stop relying on K2.6's non-thinking mode.
3. Remove temperature, top_p, n and the penalty fields from every request path, including shared wrappers.
4. Pass the complete assistant message, including reasoning_content, back into history in multi-turn and tool flows.
5. Choose a reasoning_effort level per workload and keep it fixed within a session so you keep cache hits.
6. Revisit max_completion_tokens and timeouts. The default is 131,072 output tokens, and the model always reasons.
7. Check your account tier's rate limits before moving production traffic.

Background on the model you are leaving is in our Kimi K2.7 complete guide.

FAQ

What is the Kimi K3.1 Pro API model ID?

kimi-k3.1-pro for the standard model and kimi-k3.1-pro-highspeed for the faster variant. The quickstart's base URL is https://api.moonshot.ai/v1, used with an OpenAI-compatible client.

How much does Kimi K3.1 Pro cost?

As of October 2026: $2.00 per million input tokens, $10.00 per million output tokens, $0.20 per million cached input tokens, and $2.00 or $4.00 per million for 5-minute or 1-hour cache writes. Highspeed is $4.00 / $20.00 / $0.20.

Are the Kimi K3.1 Pro weights available?

Not as of October 26, 2026. Moonshot has promised full weights by November 4, 2026. There is no Hugging Face repo yet, and Artificial Analysis lists the weights as not public.

Can I disable thinking on Kimi K3.1 Pro?

No. K3.1 Pro always reasons. Use reasoning_effort="low" to cut latency and output tokens. Do not send the K2-style thinking parameter, and do not send non-default sampling values. Moonshot says fixed parameters return an error if changed.

Is Kimi K3.1 Pro better than Claude Opus 5.5?

Not on Artificial Analysis's index, where Opus 5.5 scores 58 to K3.1 Pro's 55. K3.1 Pro is much cheaper per task ($2.11 vs $5.98) and close on Terminal-Bench 4.0. Moonshot reports 65.7% vs 66.4%, and AA measured 67% for K3.1 Pro.

When do K2 models stop working?

New API users lose access to K2 from October 31, 2026. Moonshot plans to retire all K2 models by the end of December 2026.

Sources

• Moonshot AI: Kimi K3.1 Pro launch post
• Kimi API Platform: Kimi K3.1 Pro quickstart
• Kimi API Platform: Model pricing
• Kimi API Platform: Model list and deprecations
• Kimi API Platform: Model parameter reference
• Kimi API Platform: Thinking models
• Kimi API Platform: Troubleshooting
• Kimi API Platform: Context caching
• Hugging Face: moonshotai/Kimi-K3 model card
• Artificial Analysis: Kimi K3.1 Pro
• Artificial Analysis: Claude Opus 5.5
• VentureBeat: Moonshot's Kimi K3.1 Pro flirts with open-weights leadership amid paid-influencer allegations
• The Decoder: China plans to tighten export controls on open-weight models
• Kimi Code product page
• Kimi Code documentation

If your team wants to evaluate K3.1 Pro against closed models, or self-host open-weight models like the K3 family on your own infrastructure, Codersera can help you hire vetted remote AI developers with hands-on LLM serving and evaluation experience.