Kimi K3.1 Pro: $2/$10 Pricing, Independent Score and API Limits

Moonshot AI released Kimi K3.1 Pro on October 23-24, 2026, at $2 per million input tokens and $10 per million output tokens on its API. The independent check is strong: as of October 26, Artificial Analysis gives Kimi K3.1 Pro a score of 55 on its Intelligence Index, #4 of 221 models, against 58 for Claude Opus 5.5 at max effort. Before you move a workload, though, four things deserve a look: the headline benchmarks are Moonshot's own and were run at maximum effort, the API fixes sampling and keeps thinking on, the promised open weights are not out yet, and Moonshot deliberately reduced the model's cyber capability.

What Moonshot released

The official announcement presents K3.1 Pro as the first model in a new K3.1 series and Moonshot's most capable "3T-class" open-weight model so far, with a 1-million-token context window. Moonshot's blog index dates the post October 24; The Next Web and SiliconANGLE covered it on October 23, US time.

The key facts for an API user:

• Size. Moonshot says only "3T-class". Artificial Analysis lists 2.8 trillion parameters.
• Context. The pricing docs say all K3 series models support a 1,048,576-token context window.
• Where to get it. Moonshot says it launched on Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
• Highspeed variant. A separate K3.1 Pro Highspeed model ships alongside it, which Moonshot says runs at "2-3× the speed of K3.1 Pro".
• Enterprise. Access is off by default at launch; Kimi Enterprise administrators have to enable it for their organization.
• Data retention. The blog lists Zero Data Retention for API customers but gives no further terms. If ZDR is a requirement for you, get the terms in writing before sending production data.

What it costs on the Kimi API

The announcement itself carries no prices. Moonshot's API pricing page, checked on October 26, lists these base rates:

One MTok is one million tokens. At these rates K3.1 Pro costs less than K3 for both input and output, and Highspeed costs twice the standard model on both.

Context caching is automatic, and it adds billing lines of its own:

• Cache writes are charged. The K3.1 Pro quickstart says that when an eligible prompt prefix is written to the cache, "each write incurs a one-time charge". The pricing page adds that K3 series cache writes are billed separately by TTL.
• Cache hits are billed at the cached-input price. A hit carries no extra write charge on top.
• TTL is 5m (default) or 1h. The quickstart suggests passing prompt_cache_options.ttl to pick 1h when your requests are more than five minutes apart, so an expired cache does not keep getting rewritten.

Moonshot's caching guide shows the split for kimi-k3, not K3.1 Pro: $3.00 per MTok for uncached input (a cache miss), $3.00 for a 5m write, $6.00 for a 1h write and $0.30 for a cache hit. The guide says separating out the write charge leaves the overall cost of existing requests unchanged. Those are K3 figures. If your workload leans on long cached prompts, take the K3.1 Pro write and hit rates from the current pricing page before you budget, rather than scaling K3's.

For comparison, Simon Willison puts the standard price at "the same as GPT-6.1 Sol and half the price of Claude Opus 5.5". Artificial Analysis offers a more practical number: running its index cost $2.11 per task on K3.1 Pro, using 150 million output tokens across the full run. I think that is a useful baseline: if your current model's Artificial Analysis page lists a cost per task, compare the two, then confirm on your own tasks.

The independent score, next to the competition

Artificial Analysis tested the reasoning version. On October 26 its model page gave K3.1 Pro an Intelligence Index score of 55, ranked #4 of 221 models. Ranks on that page shift as models are added, so read this as a dated snapshot. Here are nearby rows from its leaderboard on the same day, with the effort setting listed for each:

On this snapshot K3.1 Pro scores three points below Claude Opus 5.5 at max effort and two points above GPT-6 Astra at max effort. These are index points under the settings shown, not proof that two models with similar scores will do equally well on your coding work.

What Moonshot's own benchmarks claim

The announcement's tables need two qualifiers. Moonshot says that unless a table notes otherwise, its K3.1 Pro results were run at reasoning effort max, temperature = 1.0 and top-p = 0.95. And as The Next Web notes: "The figures are Moonshot's own and have not been checked independently."

Moonshot also reports gains over K3. On its internal Kimi Code Bench 3.0 long-horizon test, K3.1 Pro scores 61.2% against 55.95% for K3. Moonshot says the claim-level hallucination rate on its hard benchmark dropped from 23.4% for K3 to 8.3%. For finance, Moonshot reports state-of-the-art results on OfficeQA Pro v2 (64.4), a Frontier Finance score of 58.6, and internal finance scores where K3.1 Pro beats K3 on financial modeling (57.6 vs 50.4), financial research (70.4 vs 58.8) and business operations (37 vs 8.9).

My read: the vendor tables suggest K3.1 Pro is roughly level with Opus 5.5 on Terminal-Bench and a clear upgrade on K3. Benchmark scores like these are not a success rate you can expect on your own tasks. The independent index agrees it belongs near the top, a few points behind Opus 5.5 at max effort. Since the vendor numbers were produced at max effort, set the API to max for your own comparison, or you will be measuring a different configuration.

API limits to check before switching

The K3.1 Pro quickstart contains the details most likely to affect an existing integration:

• Thinking is always on. Thinking mode cannot be disabled; you only tune how hard the model reasons, through the top-level reasoning_effort request field. There is no non-thinking mode for simple calls.
• Effort levels are low, high and max. The docs list low, high and max, with max as the default. Simon Willison's post lists a "medium" level; follow the docs. Unless you set the field, every request runs at max.
• Sampling is fixed. The docs fix temperature=1.0, top_p=0.95, n=1, presence_penalty=0 and frequency_penalty=0, and tell you to leave those fields out of requests. If your pipeline relies on temperature 0 for repeatable extraction or classification, you cannot configure that here.
• Output length. max_completion_tokens defaults to 131,072 and can be raised to 1,048,576.
• Web search is not production-ready. Moonshot says web search is being updated and is "not recommended for production workflows in the near term". Agent workflows that depend on built-in search should wait or bring their own search tool.

Open weights are a promise for now

The announcement says: "Full model weights will be released by Nov 4, 2026." As of October 26, the moonshotai Hugging Face organization lists Kimi-K3 but no K3.1 Pro repository, and Artificial Analysis currently labels the model "Proprietary model". SiliconANGLE reports it is available through Moonshot's paid cloud services at launch. The announcement names no license for the coming weights.

If self-hosting or a specific license is the reason you are interested, wait for the model card and license before planning around it.

Reduced cyber capability, by design

For the open-weight release, Moonshot says it removed cyber-related data from both training stages while keeping general-purpose coding tasks. Its own xdev-bench-hard results (pass@3, where higher means more offensive-cyber capability) show the effect: 81.0 for Kimi K3.1 Pro against 90.5 for a Kimi version without that risk reduction, and 33.3 for K3. SiliconANGLE reports that on cyber benchmarks K3.1 Pro "significantly trails the scores of Astra and Opus 5.5".

These are benchmark scores, and they do not tell you how the model will perform on a particular security engagement. If you do authorized security research or penetration testing with a model, treat the lower cyber scores as a reason to test K3.1 Pro on your own tasks before relying on it.

Speed

On October 26, Artificial Analysis measured 49.7 output tokens per second for the standard model, #165 of 182 models, so on that measurement it is slow. Highspeed costs twice as much for a 2-3x speedup that is Moonshot's claim, so test it on your own latency-sensitive path before paying the premium.

Should you try it?

I did not test K3.1 Pro myself, so this recommendation rests on the published numbers and docs above.

• Worth a trial: agentic coding and terminal workloads where you currently pay for Opus 5.5 or GPT-6 Astra and cost matters. The independent index puts K3.1 Pro within three points of Opus 5.5 at max effort, and Simon Willison describes the price as half of Opus 5.5. Run your own evaluation at reasoning_effort max, then at high, and compare cost per completed task, not per token.
• Finance and document-QA teams: Moonshot's finance results are vendor benchmarks, several of them internal. Treat them as a reason to test, not a result.
• Hold off if you need deterministic output at temperature 0, a non-thinking mode, fast output from the standard model, production web search, or released weights with a known license.

If you used the previous generation, the Kimi K2.5 vs Claude comparison covers where the cheaper Kimi model held up and where it did not. For the OpenAI side of this comparison, see what changed in GPT-6 Astra. Moonshot's tables also use DeepSeek-V4.1-Flash as a reference row; the DeepSeek page covers that model family. I recommend checking Moonshot's Hugging Face page after November 4 before making any self-hosting plan.