Kimi K3.1 Pro Review: $2/$10 Pricing, Open Weights Due November 4 and the Claims That Do Not Match
Moonshot's new 3T-class model ranks fourth on an independent index and costs a third less than K3. The open weights are still a promise, and a paid-promotion allegation hangs over the launch.
Jump to 7 sections
Kimi K3.1 Pro is Moonshot AI's new 3-trillion-parameter model, launched October 23 and 24. It costs $2 per million input tokens and $10 per million output tokens, a third less than Kimi K3. Artificial Analysis scores it 55, fourth overall. The open weights are promised for November 4, so for now it is available only through Moonshot's apps and API.
Model Drop covers the apps, dev tools and platforms built on new models, and this review is for developers deciding whether to point a workload at Kimi this week. We checked the price list, the independent scores and the vendor claims against each other, and we flag where they disagree. One marketing question is also open, and it affects how much weight to give the buzz.
What Did Moonshot Ship?
Moonshot shipped a post-trained version of Kimi K3, released in July. According to VentureBeat's report on the launch, K3.1 Pro has about 3 trillion parameters and is available in Moonshot's apps and API. Sources differ slightly on size, since K3 is described at 2.8 trillion and Moonshot's own wording is "3T-class."
AIHubMix's launch roundup adds the specs. The model accepts image and video input, has a 1 million token context window and always reasons before it answers. Effort levels are low, high and max, with max as the default. Output can reach 1,048,576 tokens, though the default cap is 131,072 unless you raise it.
Availability has a catch. Moonshot says the weights will be released under an open license on November 4, 2026. Until then, the model is reachable only through Moonshot's apps and API, and VentureBeat notes that organizational access starts disabled and must be turned on by an administrator. That is the kind of waitlist-style detail worth knowing before you promise a team access.
Always-on reasoning has a cost. A model that thinks before every answer produces more output tokens than a model that can skip the step, and output tokens are the expensive ones. That makes the effort setting an important control, so start at low or high and only move to max when a task needs it.
How Much Does Kimi K3.1 Pro Cost?
The list price is $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens. A Highspeed variant costs twice as much on input and output. The price is a third lower than Kimi K3, so the claim that K3.1 costs the same as K3 is wrong, as AIHubMix points out.
The table below uses the prices reported by VentureBeat, in dollars per million tokens.
| Model | Input | Output | Cached input |
|---|---|---|---|
| Kimi K3.1 Pro | $2 | $10 | $0.20 |
| Kimi K3.1 Pro Highspeed | $4 | $20 | $0.20 |
| Kimi K3 | $3 | $15 | $0.30 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| DeepSeek V4.1 Flash 1017 (peak) | $0.30 | $1.20 | $0.006 |
Notice that K3.1 Pro matches Claude Sonnet 5.5 on list price. It is not a bargain next to Anthropic's mid-tier model, and it is far above DeepSeek's cheap tier. The case for Kimi rests on performance per dollar and the coming open weights, not on the lowest sticker price. Our look at the DeepSeek V4 open-weight price ceiling shows how low that floor can go.
What Do the Benchmarks Say?
The independent numbers are good, and the vendor numbers are mixed. VentureBeat reports that Artificial Analysis scores K3.1 Pro at 55 on its Intelligence Index, up from 44 for K3, ranking it fourth overall. It sits above OpenAI's GPT-6 Astra and Meta Muse Wonder, and one point below Claude Sonnet 5.5.
Other independent results show large jumps from K3 to K3.1 Pro, according to the same report.
- Terminal-Bench 4.0: 13% to 67%.
- CyberGym-E2E-AA: 58% to 95%.
- AutomationBench-AA: 58% to 73%.
- GDPval-AA: 1,534 to 1,846 Elo.
Moonshot's own comparisons are less one-sided. It reports Terminal-Bench 4.0 at 65.7 versus 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra, and SWE-Marathon at 51.7 versus 46.3 and 42.5. On Terminal-Bench Science, though, Kimi scores 38.1 against 63.3 and 68.1. VentureBeat stresses these are not controlled head-to-head tests.
Cost per task tempers the price story. Artificial Analysis puts Kimi at $1.98 per Intelligence Index task in its launch report, and $2.11 on its live page versus $5.98 for Claude Opus 5.5. That is a real saving over Opus, but it is higher than the token price alone would suggest, which suggests the model spends a lot of tokens when it reasons.
The pattern across these numbers is a model that is strong on agentic coding and automation but uneven on science and reasoning-heavy tests. Terminal-Bench Science is the clearest gap. If your work involves heavy scientific or mathematical reasoning, the vendor's own table suggests Opus 5.5 or Astra is still ahead.
What Is the Open-Weights Catch?
The catch is that nothing is open yet. Moonshot has promised weights by November 4 and an open license, but neither the files nor the license text exist today. Until they do, you cannot self-host, fine-tune or audit the model, and you should not describe it as open source in a procurement document.
Past releases show why the license matters. Some open-weight models ship with restrictions on commercial use, user counts or specific industries. Wait for the actual terms before you commit to a plan that depends on running Kimi inside your own infrastructure.
It also pays to plan for hardware. A model in the 3-trillion-parameter class is too big for a single workstation. Self-hosting means a multi-GPU cluster or a managed inference provider, and providers can take days or weeks to list a new open model. Our guide to Gemini 4 Argon pricing shows how introductory rates and access rules can change the math for platform choices.
What About the Paid-Promotion Allegation?
An X user known as Elshayib posted screenshots that he described as instructions for a Kimi promotional campaign, VentureBeat reports. One line said "Please DO NOT put on the paid partnership label," and the suggested talking points described Kimi as an open-source equivalent of Claude Opus 5.5 at a fraction of the price.
The document has not been independently authenticated. It is unclear who wrote the instructions, how many creators received them or whether any paid endorsements ran. VentureBeat found no confirmed response from Moonshot as of October 25.
Treat the allegation as unproven, and treat the enthusiasm around the launch with some care. The talking point itself is worth checking regardless of the source. The model is not open source today, and its token price is the same as Sonnet 5.5, so the "fraction of the price" claim only holds against Opus.
Reading the whole picture, the launch is a story about price pressure at the top of the market. A model near the top of an independent index for the price of a mid-tier rival, with weights coming, forces Western labs to justify their premiums. Whether it holds depends on how the weights, license and independent reruns turn out in the next week.
Who Should Use Kimi K3.1 Pro Now?
Teams with coding-agent or automation workloads should test it, because the independent scores on those tasks are strong. Teams that need self-hosting should wait for the weights and license, and teams with strict data rules should check where Moonshot's API processes data before sending anything sensitive.
A fair test is simple. Run 30 to 50 of your own tasks through Kimi, Claude Sonnet 5.5 and your current default. Record pass rate, tokens used, cost per task and latency. Our NVIDIA Nemotron 3.5 launch review uses a similar approach to separate vendor claims from results.
Keep a fallback. Because organizational access starts disabled and the weights are pending, build your integration behind an interface that lets you swap models in minutes. That protects you if access, pricing or the license changes after launch.
Data handling needs its own check. Sending prompts to a model hosted by a company in another jurisdiction raises questions about retention, logging and legal access. Read Moonshot's terms for the API and the enterprise product, and keep regulated or confidential data out of the test until your security team signs off.
The Verdict on Kimi K3.1 Pro
Kimi K3.1 Pro is a strong model at a fair price, with independent scores that back up the headline. It is not cheaper than Sonnet 5.5 on list price, it is not open yet and its launch marketing is under a cloud.
At Model Drop, we will update this review when the weights and license appear on or before November 4. Your next step is to run a small bake-off this week through the API, and to hold off on any self-hosting plan until you have read the license.
