Kimi K3.1 Pro: Pricing, Benchmarks, Open Weights
Kimi K3.1 Pro costs $2 and $10 per million tokens, a third below Kimi K3, and scores 55 on the AA index for $2.11 a task, 61% less than Claude Sonnet 5.5.
Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, a third less than Kimi K3, and scores 55 on the independent Artificial Analysis index, fourth of any model. A task costs $2.11, level with K3.
That is the verdict in one line: the rate card fell, the bill did not, and the same money now buys an 11-point better model. K3.1 Pro shares its $2 and $10 rate card with Claude Sonnet 5.5, which scores one point higher for $5.46 a task, and with Gemini 4 Argon’s introductory price, which ends on November 30. For a model within three points of the top of the index, nothing else costs as little per task, and Moonshot says the weights ship by November 4.
Moonshot released Kimi K3.1 Pro on October 23, 2026, as a post-trained version of Kimi K3 with “substantially larger post-training compute”. It is live on the Kimi API, Kimi.ai, the Kimi apps and Kimi Code.
| Kimi K3.1 Pro | |
|---|---|
| Input, per Mtok | $2.00 |
| Output, per Mtok | $10.00 |
| Cached input (cache hit) | $0.20 |
| Cache write, 5-minute / 1-hour | $2.00 / $4.00 |
| Highspeed variant (2-3x faster) | $4.00 / $20.00 |
| API ids | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
| Available | Kimi API, Kimi.ai, Kimi apps, Kimi Code, October 23 |
| Weights | Promised by November 4, 2026 |
| Size | 2.8T total, 104B active (mixture of experts) |
| Context window | 1,048,576 tokens |
| AA Intelligence Index | 55 (fourth of any model) |
| AA cost per index task | $2.11 |
How much does Kimi K3.1 Pro cost?
$2.00 per million input tokens, $0.20 per million cached input tokens and $10.00 per million output tokens, according to the Kimi API pricing page. Every rate is exactly two thirds of Kimi K3’s $3.00, $0.30 and $15.00, and K3 stays on the same page at its old price, so nothing forces a migration.
Two lines on the rate card are easy to miss. Cache writes bill separately: $2.00 per million tokens to hold a prefix for five minutes and $4.00 for an hour, while a cache hit costs only the $0.20 read. And the Highspeed variant, which Moonshot says runs two to three times faster, costs exactly double: $4.00 and $20.00. That makes Highspeed the same price as Claude Opus 5.5.
The same page carries a notice for anyone on older Kimi models: the K2 series, including K2.7 Code, closes to newly registered users on October 31, 2026, and Moonshot plans to retire all K2 models by the end of December. The migration link points to K3.1 Pro.
Is Kimi K3.1 Pro cheaper than Kimi K3?
Per token, yes. Per task, no. Artificial Analysis measures $2.11 per Intelligence Index task for K3.1 Pro against $2.00 for Kimi K3. Its launch write-up put the figure at $1.98, “roughly equal to Kimi K3”; the live model page now reads $2.11.
| Step | Change | Running total |
|---|---|---|
| Kimi K3 (Measured, $3 / $15 rates, scores 44) | $2.00 | $2.00 |
| One-third rate cut (Every rate x 2/3) | -$0.67 | $2.00 |
| K3 workload, new rates (Modeled) | $1.33 | $1.33 |
| Heavier workload (+14% output tokens, longer runs) | +$0.78 | $2.11 |
| Kimi K3.1 Pro (Measured, scores 55) | $2.11 | $2.11 |
The output tokens alone do not explain the step. Artificial Analysis counts 55,000 output tokens per task for K3.1 Pro against 48,000 for K3, a 14% rise. If the whole bill scaled with output tokens, the new rates would land near $1.52, not $2.11. The rest is spent on input and cache traffic, and the likeliest cause is that the model now finishes long agent tasks K3 gave up on: its Terminal-Bench 4.0 score in the Artificial Analysis run rose from 13% to 67%. A model that keeps working reads more context.
So the honest reading of the price cut is that Moonshot made the model better at the same price per task, not cheaper at the same quality. If your workload is short and single-turn, you will see most of the one-third saving. If it is an agent loop, budget for K3’s bill.
What does a task cost on Kimi K3.1 Pro compared with other models?
$2.11 per index task buys 55 points. Read off each model’s Artificial Analysis page on October 24, 2026, at its highest-scoring effort setting:
| Model | AA index | Cost per index task | Output tokens | Price per Mtok (in / out) |
|---|---|---|---|---|
| Claude Opus 5.5 (max) | 58 | $5.98 | 260M | $4.00 / $20.00 |
| Claude Sonnet 5.5 (max) | 56 | $5.46 | 420M | $2.00 / $10.00 |
| Kimi K3.1 Pro | 55 | $2.11 | 150M | $2.00 / $10.00 |
| Gemini 4 Argon (high) | 53 | $1.99 | 110M | $2.00 / $10.00 (intro) |
| GPT-6 Astra (max) | 53 | $3.26 | 60M | $10.00 / $50.00 |
| GLM-5.3 (max) | 45 | $2.01 | 210M | $1.40 / $4.40 |
| DeepSeek V4.1 Flash 1017 (max) | 45 | $0.26 | 210M | $0.30 / $1.20 (peak) |
| Kimi K3 (max) | 44 | $2.00 | 160M | $3.00 / $15.00 |
Sources: Claude Opus 5.5, Claude Sonnet 5.5, Kimi K3.1 Pro, Gemini 4 Argon, GPT-6 Astra, GLM-5.3, DeepSeek V4.1 Flash 1017 and Kimi K3.
The three $2 and $10 models make the cleanest comparison on the page, because the rate card is identical and only the tokens differ. Sonnet 5.5 burns 420 million output tokens to run the index, K3.1 Pro 150 million and Argon 110 million. That is why the per-task costs spread from $1.99 to $5.46 on one price.
The Artificial Analysis totals and the launch article’s per-task counts do not line up neatly: the model pages list 150 million output tokens for K3.1 Pro and 160 million for K3, while the article reports 14% more tokens per task for K3.1 Pro. The per-task figures above are the ones Artificial Analysis bills against, so they are the ones used here.
Is Kimi K3.1 Pro better than Claude Opus 5.5?
Not on the independent index: Opus 5.5 scores 58 to K3.1 Pro’s 55, for nearly three times the cost per task. Moonshot’s own benchmark table has K3.1 Pro close behind on coding. Every score in this table is vendor-reported by Moonshot in its launch post and not independently reproduced.
| Benchmark (Moonshot-reported) | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% | Not reported |
| BenchCAD Voxel IoU, with tools | 92.7 | 96.2 | 95.9 | 92.6 |
| BenchCAD Voxel IoU, without tools | 72.8 | 73.0 | Not reported | 60.6 |
| Hallucination rate, no tools (lower is better) | 5.4 | 6.8 | 4.2 | 11.4 |
| Hallucination rate, hard with tools | 8.3 | 15.0 | 5.5 | 23.3 |
Two independent reads point the same way. On the GDPval-AA leaderboard, Artificial Analysis’s run of professional-deliverable tasks, K3.1 Pro rates 1,846 Elo, third behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. And on Terminal-Bench 4.0 in Artificial Analysis’s own harness it scores 67%, which the launch write-up calls the highest of any model it has tested.
Where it falls short is breadth. The hallucination rows are Moonshot’s own test, and GPT-6 Astra beats it on both. The Artificial Analysis Omniscience hallucination rate is 44%, down from K3’s 53% but still high. For an agent that has to be right without a check, Opus 5.5 is still the safer default; for an agent whose output gets tested, the price gap is hard to ignore.
Kimi K3.1 Pro vs Gemini 4 Argon: which is cheaper?
Today they are a tie: both cost $2 and $10 per million tokens, and Argon’s $1.99 per index task is twelve cents under K3.1 Pro’s $2.11 for two points less. The difference is the calendar. Argon’s $2 and $10 is an introductory price that Google’s pricing page runs through November 30, 2026, after which it becomes $4 and $20, as covered in the Gemini 4 Argon breakdown. On a straight doubling of every rate, an Argon task costs about $4 from December 1.
Kimi’s $2 and $10 is its standard rate, with no stated expiry. If you are choosing a model to build on past November, K3.1 Pro is the cheaper of the two on a per-task basis and the higher scorer. If you need Google’s multimodal stack or a US provider, Argon is the one to test, and it stays cheap for five more weeks.
When are the Kimi K3.1 Pro open weights released?
By November 4, 2026, according to Moonshot’s launch post: “The full model weights will be released by Nov 4.” Until then Artificial Analysis lists the model as proprietary. Its launch write-up says that once the weights are out, K3.1 Pro will be the highest-scoring open-weight model, narrowing the gap between open and closed models on its index from 12 points to 3.
Moonshot’s track record on this is good: it promised Kimi K3’s weights by July 27 and shipped them a day early, as the Kimi K3 breakdown records. The licence for K3.1 Pro is not stated in the launch post.
The release comes with a deliberate limit. Moonshot says it removed cyber-related data from mid-training and filtered cyber-related reinforcement-learning environments for the open-weight model, so that its offensive cyber capability stays below what frontier closed models already offer. On its xdev-bench-hard exploit test, Moonshot reports 81.0% for K3.1 Pro, against 90.5% for an internal checkpoint without those measures and 33.3% for K3. Those numbers are vendor-reported.
Should you use Kimi K3.1 Pro?
Four cases:
- You run Kimi K3 today. Switch. You pay about the same per task for an 11-point better model, and the K2 line you might fall back to is being retired.
- You run Claude Sonnet 5.5 for long agent loops. Test K3.1 Pro on your own tasks. On the independent index it is one point behind at 39% of the cost per task, though Sonnet’s margin on breadth and reliability is real.
- You need the top score and the price is secondary. Stay on Claude Opus 5.5, which still leads the index and Moonshot’s own Terminal-Bench table.
- You want cheap volume, not the frontier. Use DeepSeek’s October build of V4.1 Flash, which scores 45 for $0.26 a task, if your procurement rules allow a Chinese model.
To price your own task shape, run it through the AI model comparison tool. Every current Moonshot model is on the AI model tracker, and the rest of the month’s launches are in new AI models released in October 2026. The company behind it is profiled in Moonshot AI explained.
How to get started with Kimi K3.1 Pro
- Top up the Kimi API. Create a key on the Kimi API Platform and call
kimi-k3.1-prothrough its OpenAI-compatible endpoint. The quickstart covers tool calling, structured output and video input. - Use Moonshot’s settings. Its benchmarks ran at reasoning effort
max, temperature 1.0 and top-p 0.95. Lower effort cuts tokens, and with them the cost per task. - Plan your cache. Repeated prefixes are cached automatically for five minutes. If your agent pauses longer, pay the $4.00 one-hour write once rather than the $2.00 input rate on every call.
- Reach for Highspeed only when latency matters. At $4 and $20 it costs the same as Claude Opus 5.5.
Frequently asked questions
- How much does Kimi K3.1 Pro cost?
- $2.00 per million input tokens, $0.20 per million cached input tokens and $10.00 per million output tokens on the Kimi API, a third below Kimi K3. Cache writes cost $2.00 per million tokens for a five-minute hold and $4.00 for one hour. The Highspeed variant costs $4.00 and $20.00.
- Is Kimi K3.1 Pro cheaper than Kimi K3?
- Per token it is a third cheaper. Per task it is not: Artificial Analysis measures $2.11 per Intelligence Index task against $2.00 for Kimi K3, because K3.1 Pro spends about 14% more output tokens and runs longer. It also scores 55 to K3 at 44.
- Is Kimi K3.1 Pro open source?
- Not yet. Moonshot says the full weights will be released by November 4, 2026, and has not stated the licence. Until then Artificial Analysis lists the model as proprietary. Kimi K3, its predecessor, has had open weights since July 2026.
- How does Kimi K3.1 Pro compare with Claude Opus 5.5?
- Opus 5.5 scores 58 on the Artificial Analysis Intelligence Index to Kimi K3.1 Pro at 55, and costs $5.98 per index task against $2.11. On Moonshot-reported Terminal-Bench 4.0 the two are close, 66.4% for Opus 5.5 and 65.7% for K3.1 Pro.
- What is Kimi K3.1 Pro Highspeed?
- A faster variant of the same model, served as kimi-k3.1-pro-highspeed, which Moonshot says runs two to three times faster. It costs double the standard rate: $4.00 per million input tokens and $20.00 per million output tokens.
Sources
Moonshot AI (2026). Kimi K3.1 Pro. Kimi tech blog, October 23, 2026: post-trained K3, Highspeed variant, vendor-reported benchmarks, cyber risk-reduction, weights by November 4. https://www.kimi.ai/blog/kimi-k3-1-pro Verified 2026-10-24.
Moonshot AI (2026). Model pricing. Kimi API Platform documentation: kimi-k3.1-pro, kimi-k3.1-pro-highspeed and kimi-k3 rates, cache-write pricing, K2 series retirement notice. https://platform.kimi.ai/docs/pricing Verified 2026-10-24.
Moonshot AI (2026). Kimi K3.1 Pro quickstart. Kimi API Platform documentation. https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart Verified 2026-10-24.
Artificial Analysis (2026). Kimi K3.1 Pro launches at 4th on the Intelligence Index. Artificial Analysis article, October 23, 2026: per-evaluation gains over K3, tokens per task, open-weights outlook. https://artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks Read 2026-10-24.
Artificial Analysis (2026). Kimi K3.1 Pro, Kimi K3, Claude Opus 5.5, Claude Sonnet 5.5, Gemini 4 Argon, GPT-6 Astra, GLM-5.3 and DeepSeek V4.1 Flash 1017 model pages, and the GDPval-AA leaderboard. Independent benchmark index. https://artificialanalysis.ai/models/kimi-k3-1-pro and https://artificialanalysis.ai/evaluations/gdpval-aa Read 2026-10-24.