Kimi K3.1 Pro: Benchmarks, Price Drop and How to Use It
Moonshot AI released Kimi K3.1 Pro on October 23, 2026. It scores 55 on the Artificial Analysis Intelligence Index — fourth overall and 3 points behind Claude Opus 5.5 — while API prices fall 33% to $2 per million input tokens and $10 per million output tokens. It is available today on Kimi.com, Kimi Work, Kimi Code and the Kimi API; the model weights are due by November 4, 2026.
Last checked: October 24, 2026. Sources: Moonshot's official K3.1 Pro tech blog, the Kimi API pricing page and Artificial Analysis.
One thing to get out of the way: aiwith does not host K3.1 Pro yet. The model on aiwith is still Kimi K3. If you want the newer model today, use Moonshot's own products or API. If you want a Kimi model with a 1M-token context inside a single subscription alongside DeepSeek, GPT and Claude, Kimi K3 is available now — details at the end.
What changed in K3.1 Pro?
K3.1 Pro is not a new architecture. It is a post-trained version of Kimi K3 with substantially more post-training compute, including large-scale reinforcement learning across hundreds of downstream domains. Moonshot calls it its most capable 3T-class open-weight model to date.
The headline specs:
- 1,048,576-token context window, same as K3.
- Text and images as input. No audio or video.
- 2.8 trillion total parameters with 104 billion active, according to The Decoder — the same dimensions as K3. Moonshot describes the model as "3T-class."
- A Highspeed variant that runs two to three times faster.
The gains are concentrated in agentic work, coding, 3D creation and financial analysis rather than chat quality. Moonshot's own framing is that K3.1 Pro "matches or approaches the strongest proprietary models" on coding, agentic and knowledge benchmarks.
Benchmarks: where K3.1 Pro lands
Artificial Analysis puts K3.1 Pro at 55 points on its Intelligence Index, up 11 points from K3 and fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). If the weights ship as planned, it would be the highest-scoring open-weights model available and shrink the open-versus-proprietary gap from 12 points to 3.
Selected scores, all with reasoning effort set to max:
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | DeepSeek V4.1 Flash |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 42.3% |
| SWE Marathon v1.1 | 51.7% | 46.3% | 40.0% |
| DeepSWE v1.1 | 75.5% | 74.2% | 77.0% |
| BenchCAD (with tools) | 92.7% | 96.2% | 81.6% |
| HLE-Full (with tools) | 70.0% | 67.7% | 58.8% |
| BrowseComp | 92.1% | 90.8% | — |
| OfficeQA Pro v2 | 64.4% | 62.2% | 40.4% |
| Automation Bench-AA | 73.3% | 69.5% | 66.6% |
| GDPval-AA v2.1 (Elo) | 1,846 | 1,866 | 1,746 |
| GPQA-Diamond | 94.2% | 93.9% | 92.2% |
| ExploitBench | 74.7% | 91.0% | — |
Two numbers stand out beyond the table. On GDPval-AA, a benchmark for economically valuable real-world work, K3.1 Pro gains 312 Elo over K3 and lands within 20 Elo of Claude Opus 5.5. On Moonshot's internal Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4% with K3 to 8.3% — though Moonshot explicitly says this is a stress test built from error-prone prompts, not a measure of everyday error rates.
The caveats are the usual ones. Moonshot's benchmark table is vendor-published. Artificial Analysis supplies the independent Intelligence Index and GDPval numbers. Neither measures your workload. Treat the ranking as a reason to test, not a reason to migrate.
Pricing: 33% cheaper than K3
The price cut is the most immediately useful change. Moonshot kept the model the same size and cut token prices by a third:
| Model | Input | Output | Cache hit |
|---|---|---|---|
| kimi-k3.1-pro | $2.00 / MTok | $10.00 / MTok | $0.20 / MTok |
| kimi-k3.1-pro-highspeed | $4.00 / MTok | $20.00 / MTok | $0.20 / MTok |
| kimi-k3 (previous) | $3.00 / MTok | $15.00 / MTok | $0.30 / MTok |
Cache writes cost $2 per MTok for K3.1 Pro, or $4 for one-hour TTL writes. The API enables context caching automatically with a five-minute default TTL, which matters for agent loops that resend the same prompt prefix.
Per completed task the saving is smaller than the headline: Artificial Analysis measures $1.98 per Intelligence Index task for K3.1 Pro versus $2.00 for K3, because the new model generates about 14% more output tokens. Cheaper tokens, more of them, roughly the same bill — but 11 index points higher.
Highspeed costs exactly double for two to three times the throughput. If you are running an interactive coding agent, that trade is usually worth it; for batch jobs it is not.
How do I use Kimi K3.1 Pro?
- Kimi.com and the mobile apps: available now on iOS, Android and HarmonyOS.
- Kimi Work: desktop app version 3.1.0 or later, Windows and Apple silicon Macs.
- Kimi Code: run the
/modelcommand and select Kimi K3.1 Pro. - Kimi API: use the model IDs
kimi-k3.1-proorkimi-k3.1-pro-highspeed. The API platform also offers Zero Data Retention for API customers. - Kimi Enterprise: administrators can enable K3.1 Pro, but it is off by default at launch.
- Self-hosting: not yet. Moonshot says full weights will be released by November 4, 2026. Wait for the license and hardware guidance before planning an on-prem deployment.
One launch detail worth knowing: Moonshot removed cyber-related data from both mid-training and reinforcement learning before releasing K3.1 Pro. On its internal xdev-bench-hard vulnerability benchmark, the uncleaned checkpoint solves 90.5% of tasks and the cleaned model solves 81.0%. That is still far above K3's 33.3%, and below DeepSeek V4.1 Flash's 85.7% on the same test. Moonshot says it deliberately keeps the open-weights release below the offensive-cyber frontier.
Can I use K3.1 Pro on aiwith or Router?
Not yet on aiwith. aiwith's Kimi line-up is still Kimi K3, a 1M-token model with reasoning always on. K3.1 Pro has not been added, so a subscription does not unlock it. If you want a Kimi model inside aiwith today:
- Kimi K3 — the nearest match: 1M-token context, strong long-document and whole-repo work.
- DeepSeek V4.1 Flash — the cheaper option for single-file fixes, scripts and high-volume chat.
- GPT and Claude — frontier proprietary models when the Kimi/DeepSeek tier is not enough.
On Router, K3.1 Pro is not listed either. Router carries kimi/kimi-k3 today, with the same 1M-token context and tool calling. When Moonshot's model appears in Router's model list, it will be the authoritative place to check. Until then, use the Kimi API directly for K3.1 Pro.
Is it worth switching?
It depends on what you are switching from.
From Kimi K3: yes, if you do agentic work. You get 11 index points, a 312-Elo gain on GDPval, and cheaper input tokens for the same price per task. There is no migration risk beyond a model ID change, and K3 stays available. If you only use Kimi for long-context reading, the 1M window has not changed and the upgrade is optional.
From Claude Opus 5.5 or GPT-6 Astra: not on benchmarks alone. K3.1 Pro is within 3 index points and beats both on OfficeQA Pro v2, HLE-Full and DeepSWE, but trails badly on ExploitBench (74.7 vs 91.0) and slightly on Terminal-Bench. If your work is finance, document analysis or cost-sensitive agent loops, it is a serious candidate. If your work depends on the top end of coding or security tooling, test before you move.
From DeepSeek V4.1 Flash: they are different tools. DeepSeek is dramatically cheaper for high-volume, short-turn work and still wins DeepSWE. K3.1 Pro wins long-horizon coding, browsing, office and finance tasks. Running both is usually cheaper than forcing one to do everything.
If you are waiting for open weights: set a reminder for November 4. Until the files and license are published, "open" is a plan, not a fact.
The practical next step is a two-hour test on your own material: one long-horizon coding or document task, one agent task with tools, and one task you already know the right answer to. Compare K3.1 Pro against whatever you use today on correctness, retries and total tokens. A 3-point index gap will not survive contact with your actual workload if the model mishandles your file formats.