Kimi K3.1 Pro: Opus-Level Charts, Mixed Real-World Results

Moonshot's Kimi K3.1 Pro matches Opus 5.5 on its own benchmarks at half the price. Independent testers found a slower, wordier model. What to believe.

Disclosure: this article is based on Moonshot AI’s published benchmarks and pricing, plus independent tests reported by third parties. We have not run Kimi K3.1 Pro on our own workloads. Benchmark figures are attributed, and most of the favourable ones come from Moonshot.

Moonshot AI released Kimi K3.1 Pro on October 23, 2026. It is a post-trained version of Kimi K3, the 2.8-trillion-parameter open-weight model Moonshot launched in July, and on Moonshot’s own charts it lands within a point of Claude Opus 5.5 on Terminal-Bench 4.0 at half the API price: $2 per million input tokens and $10 per million output.

The honest summary is that the headline and the independent picture do not match. Outside testers found Kimi K3.1 Pro slow, unusually verbose and behind Opus 5.5 on several real tasks, and Artificial Analysis cut its ranking sharply after an index update. It is a capable open-weight coding model at an attractive price. It is not, on current evidence, an Opus replacement.

Table of Contents

Kimi K3.1 Pro climbing rope on a mountain route
A benchmark chart is the route on the map. Independent tests of Kimi K3.1 Pro are the climb, and the climb was harder than the map suggested.

What Kimi K3.1 Pro Is

Kimi K3.1 Pro is Moonshot’s coding-focused refresh of Kimi K3. In its launch post, Moonshot described it as a post-trained version of K3 with substantially larger post-training compute. The base architecture is unchanged; what changed is how much reinforcement learning and fine-tuning went on top.

It is available on Kimi.ai, Kimi Work, Kimi Code and the API, under the model names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Vercel added both to its AI Gateway on launch day, which makes it easy to try without a Moonshot account.

The context window is one million tokens. Like its parent, the model routes each token to 16 of 896 experts and activates about 104 billion parameters per token, according to BuildMVPFast’s summary of the model details.

The pitch is simple: Opus-class coding at Sonnet-class prices, with open weights to follow. The first two parts are what everyone noticed. The third turned out to be complicated.

The Kimi K3 Foundation

To understand Kimi K3.1 Pro you need the model it is built on. Moonshot released Kimi K3 on July 16, 2026, as a 2.8-trillion-parameter mixture of experts, then the largest open-weight model by parameter count. Decrypt noted it roughly doubled the size of the nearest open competitor; DeepSeek’s V4 Pro topped out at 1.6 trillion.

K3 introduced two architectural ideas Moonshot calls Kimi Delta Attention and Attention Residuals, and shipped with MXFP4 weights, a 4-bit format that makes a model this large slightly less impossible to serve. The weights went up on Hugging Face on July 27.

K3 was priced at $3 per million input tokens and $15 per million output, with cached input at $0.30. The Decoder called it a sign of the end of super-cheap Chinese AI, because it priced exactly like Anthropic’s Sonnet 5 at the time. Independent testing from Artificial Analysis still put K3 behind Claude Fable 5 and GPT-5.6 Sol, but ahead of most other models.

In Meta’s October comparison table for Muse Spark Open, Kimi K3 scored 85.0 on Terminal-Bench 2.1, well ahead of Meta’s model. That reputation as a strong coder is the platform K3.1 Pro launched from.

Kimi K3.1 Pro Pricing and the Highspeed Option

Model Input / output per 1M Cached input per 1M
Kimi K3.1 Pro $2.00 / $10.00 $0.20
Kimi K3.1 Pro Highspeed $4.00 / $20.00 $0.20
Claude Opus 5.5 $4.00 / $20.00 $0.20
GPT-6 Astra $10.00 / $50.00 $1.00
DeepSeek V4.1 Flash (1017) $0.15 / $0.60 $0.003

The standard model is cheaper than K3 was, which is a reversal of the price direction K3 had set. It matches Claude Sonnet 5.5 and GPT-6.1 Sol on list price.

The Highspeed version is marketed as two to three times faster and costs double, landing exactly on Opus 5.5’s price. That is an awkward position: if you are paying Opus prices, the comparison is no longer favourable on cost, only on openness.

Subscribers ran into another problem early. BuildMVPFast reported that users hit five-hour quota limits quickly, and Moonshot apologised on October 25 and gave a one-time quota reset worth 25% of the monthly allowance. Early demand outpacing capacity is common for launches, but it is worth knowing if you plan to rely on Moonshot’s own plans.

Kimi K3.1 Pro Benchmarks, by Moonshot

These are Moonshot’s numbers at max effort, using the Claude Code scaffold, as reported by BuildMVPFast.

Benchmark Kimi K3.1 Pro Claude Opus 5.5 GPT-6 Astra DeepSeek V4.1 Flash 1017
Terminal-Bench 4.0 65.7 66.4 58.2 42.3
SWE Marathon v1.1 51.7 46.3 42.5 40.0
DeepSWE v1.1 75.5 74.2 74.1 77.0
BenchCAD, with tools 92.7 96.2 95.9 81.6

On these numbers, Kimi K3.1 Pro is genuinely competitive with Opus 5.5: slightly behind on Terminal-Bench and BenchCAD, ahead on SWE Marathon and DeepSWE. Two details deserve attention. Moonshot’s footnotes mark proprietary scores as with fallback, a setting that can affect results. And DeepSeek’s much cheaper Flash model edges everyone on DeepSWE, which says something about how narrow some of these margins are.

What Outside Testers Found

This is where the story changes. BuildMVPFast collected several independent tests, and the pattern is consistent.

  • Planted-bug test: Paweł Huryn planted 105 bugs across two repositories. DeepSeek V4.1 Flash 1017 scored 42.7, Muse Wonder Full 42, Opus 5.5 41.7 and Kimi K3.1 Pro 36.
  • VulcanBench: Morgan Linton’s low-effort run solved none of four tasks, at 26 minutes per task.
  • FrontierML: Induction Labs found it the best open-weight model at training other AI models, but far behind Opus 5.5 at every budget.
  • Artificial Analysis: initially ranked it 55, fourth overall according to a Hacker News comment, then 46, twenty-second of 62, after an index update.

Each test has its own limits, and individual testers are not a representative sample. But when several independent tests point the same way and the main independent index drops nine points on an update, the vendor’s charts deserve less weight than usual.

Simon Willison’s quick check is a useful counterpoint. He reported the model identifying a raw-weights puzzle after about ten minutes and nearly 20,000 reasoning tokens. It can solve hard problems. It just takes a long time and a lot of tokens to do it.

Speed, Verbosity and the Real Cost per Task

Two numbers from Artificial Analysis explain most of the gap between list price and real cost. Kimi K3.1 Pro produced 110 million output tokens on the index run, against a median of 73 million. And its output speed was 30.8 tokens per second, 194th of 196 models measured.

Put those together and the $2 and $10 list price looks different. A model that writes half again as many tokens as its peers pays half again as much in output charges for the same task. Artificial Analysis put the cost at $3.68 per index task. And at about 31 tokens per second, a long answer takes minutes.

BuildMVPFast also noted that Artificial Analysis described the model’s personality change as the biggest it had seen versus a direct predecessor. In practice, that means prompts tuned for K3 may behave quite differently on K3.1 Pro. Retest before switching.

The Highspeed variant addresses speed, at double the price. Whether it closes the gap depends on whether the extra speed comes with the same verbosity.

The Licence and the Missing Weights

Kimi K3 shipped under the Kimi K3 License, which BuildMVPFast describes as MIT-like with two conditions:

  • Model-as-a-service: revenue above $20 million over 12 months from offering the model as a service requires a separate agreement with Moonshot.
  • Large products: products with over 100 million monthly users or $20 million in monthly revenue must display “Kimi K3” in the user interface.

Internal use is exempt from both. For most companies those thresholds will never apply, so in practice the licence behaves like MIT.

The open-weight promise for K3.1 Pro specifically is less settled. Moonshot promised the weights by November 4, but BuildMVPFast reported they were not on Hugging Face when the author checked that day, and the page does not state a separate licence for K3.1 Pro. Until the weights appear, Kimi K3.1 Pro is effectively an API model with an open-weight parent.

Hardware is the other practical limit. At 2.8 trillion parameters, very few teams will self-host it even when the weights are available.

Safety Changes in the Open Release

Moonshot made an unusual disclosure about cyber capabilities. It removed cyber-related training data before the open release. On its in-house exploit test, the uncut checkpoint solved 90.5% and the final model 81.0%.

That trade is worth understanding. An open-weight model cannot be recalled once released, so labs increasingly reduce offensive security capability before publishing weights. The cost is a less capable model for legitimate security work; the benefit is less help for attackers who download it. Moonshot chose to give up about nine points on its own test.

Kimi K3.1 Pro vs Opus 5.5 and DeepSeek Flash

Question Kimi K3.1 Pro Claude Opus 5.5 DeepSeek V4.1 Flash
List price per 1M $2 / $10 $4 / $20 $0.15 / $0.60 off-peak
Vendor coding charts Near Opus Leader on several Strong on DeepSWE
Independent tests Mixed, often behind Strong Strong for the price
Speed Very slow at standard tier Moderate Fast
Open weights Promised, not yet seen No Yes, MIT

Read that way, the comparison favours the other two for most buyers. If you want the strongest coding results and can pay, Opus 5.5 has the better independent record. If you want an open model at the lowest cost, DeepSeek’s Flash model beat Kimi K3.1 Pro on the planted-bug test at a fraction of the price.

Kimi’s case is narrower: a large open-weight model, eventually, with strong charts on long-horizon coding tasks such as SWE Marathon. If those are your tasks and the weights arrive, it is worth a test.

How to Test It Fairly

Given the gap between charts and outside results, run your own comparison before switching anything.

Pick twenty real coding tasks with tests that decide pass or fail. Run them through Kimi K3.1 Pro, your current model and one cheap alternative. Record four numbers: pass rate, total output tokens, wall-clock time and cost. The output token count matters most here, because verbosity is where the list price misleads.

Use the same scaffold for every model. Vendor benchmarks often use the scaffold that suits their model best, and scaffold choice can swing results by several points.

Check latency in your real workflow, not in a batch. A model that produces 31 tokens per second may be fine for overnight jobs and painful for interactive coding.

Finally, retest your prompts. If the personality change from K3 is as large as reported, instructions that worked before may need rewriting.

What Kimi K3.1 Pro Says About Open Models

The release captures a tension running through the open-weight market. Chinese labs have pushed open models close to the closed frontier on benchmark charts, and their prices have forced Western labs to cut. But charts and real-world performance are drifting apart, as labs tune hard for the tests everyone watches.

Independent evaluation is the correction. Artificial Analysis updating its index and dropping the model nine points is exactly what should happen when a benchmark suite stops reflecting real capability. Practitioners publishing their own tests, even small ones, add another check.

For buyers, the lesson is not to distrust open models. It is to distrust single numbers. Look for agreement between vendor charts, independent indices and hands-on tests. When they agree, act. When they diverge, test yourself.

Where Kimi K3.1 Pro Could Still Win

None of this means the model has no place. Its strongest chart result is SWE Marathon, a benchmark of long, multi-step software tasks, where Moonshot reports Kimi K3.1 Pro ahead of both Opus 5.5 and GPT-6 Astra. If your work is dominated by long batch jobs where nobody is waiting on the answer, its slowness matters less and its price matters more.

Research on model training is another niche. Induction Labs found it the best open-weight model for training other AI models, even if it trailed Opus. Teams that want an open model for that kind of experimentation have few stronger options.

And there is the open-weight question itself. If the weights for Kimi K3.1 Pro arrive under the same terms as K3, organisations that need to run a frontier-scale coding model on their own hardware will have one more choice. That option has value even for teams who mostly use closed APIs, because it caps how far any single vendor can push prices.

Who Should Use What

If you are Start with Why Link
Wanting the strongest coding results Claude Opus 5.5 Better independent record See Opus 5.5
Wanting the cheapest capable open model DeepSeek V4.1 Flash Strong tests at a fraction of the price See specs
Running long-horizon coding batch jobs Kimi K3.1 Pro trial Strong SWE Marathon chart, speed matters less Read the review
Needing open weights today Kimi K3 or another open model K3.1 Pro weights not yet confirmed See Kimi K3

Frequently Asked Questions About Kimi K3.1 Pro

When was Kimi K3.1 Pro released?
Moonshot AI released it on October 23, 2026, as a post-trained version of Kimi K3, which launched in July.

How much does Kimi K3.1 Pro cost?
$2 per million input tokens, $10 output and $0.20 cached. The Highspeed version costs $4 and $20, the same as Claude Opus 5.5.

Is it as good as Claude Opus 5.5?
On Moonshot’s charts it is close. Independent tests were less favourable, and Artificial Analysis lowered its ranking from 55 to 46 after an index update.

Why is it slow?
Artificial Analysis measured about 31 tokens per second, near the bottom of its list, and found it writes far more tokens than most models. The Highspeed tier is marketed as two to three times faster.

Are the weights available?
Moonshot promised them by November 4, but they were not visible on Hugging Face when one reviewer checked. Kimi K3’s weights have been public since July 27.

What licence applies?
Kimi K3 uses an MIT-like licence with conditions for large model-as-a-service businesses and very large products. Internal use is exempt. A separate licence for K3.1 Pro had not been stated.

Verdict

Treat Kimi K3.1 Pro as an interesting option to test, not a proven Opus alternative. Its charts are strong and its list price is half of Opus, but independent results, speed and verbosity all cut against the headline.

The honest caveat is that nearly every favourable number comes from Moonshot, under settings it chose, and the open weights had not appeared when promised. Some outside tests are small and may not be representative either.

Next step: run twenty of your own coding tasks through Kimi K3.1 Pro, Opus 5.5 and DeepSeek’s Flash model in the same scaffold, and compare cost per passing task rather than price per token.

Sources and Further Reading

On this site: compare with Claude Opus 5.5, see the cheap alternative in DeepSeek V4.1 Flash, survey the field in open-weight models this November, and pick tools from AI coding assistants in 2026.