Moonshot AI released Kimi K3.1 Pro on October 23, 2026, and the first independent scoreboard put it fourth overall. The Decoder reports that the model scored 55 points on the Artificial Analysis Intelligence Index, an 11-point jump over its predecessor Kimi K3. Only Claude Opus 5.5, Claude Sonnet 5.5 and Meta’s Muse Wonder Full ranked higher.
Moonshot frames the release as the opening move of a new product line. In its announcement, the company said: “We’re introducing Kimi K3.1 Pro, the first model in our new K3.1 series and our most capable 3T-class open-weight model to date, with a 1-million-token context window.” A faster sibling, Kimi K3.1 Pro Highspeed, shipped the same day.
This analysis covers what Moonshot AI confirmed, what only third-party outlets have reported, and what the launch means for developers choosing between closed frontier APIs and open-weight options. If you track self-hosted models, our local LLM ranking of Muse Spark, Beam and DeepSeek-V4.1-Flash gives the field Kimi now enters.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Moonshot AI Announced on October 23
The official materials identify four things. Kimi K3.1 Pro is the first model in the K3.1 series. It belongs to the 3T-class of open-weight models. It carries a 1-million-token context window. And it targets coding, 3D creation, financial research and modeling, and long-horizon agent work.
Moonshot also made a claim about direction of travel: “It delivers substantial improvements across coding, 3D creation, financial research, and other agentic capabilities.” The company did not publish a price in the materials available to us, and it did not confirm a full weight release date there either. Both items appear below as reported by outside outlets, not as settled facts.
Unite.AI describes K3.1 Pro as a post-trained version of Kimi K3, built with substantially larger post-training compute. The Decoder agrees with that lineage. OpenRouter, which added the model to its catalog on launch day, described it in nearly identical terms: “A 1M context, multimodal open-weight model, post-trained from Kimi K3 for long-horizon coding and agentic work.” You can see the listing on OpenRouter’s Kimi K3.1 Pro page.
Kimi K3.1 Pro Highspeed and Where Both Models Run
Kimi K3.1 Pro Highspeed is the second product in the launch. Per Unite.AI and The Decoder, Moonshot says it runs two to three times faster than K3.1 Pro, with the same model capabilities. That makes it a latency tier, not a smaller or weaker model.
Where the models are available
Moonshot stated that “Kimi K3.1 Pro and K3.1 Pro Highspeed are available today on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.” Unite.AI adds detail from the announcement. Mobile apps cover iOS, Android and HarmonyOS. The Kimi Work desktop app needs version 3.1.0 or later on Windows and Apple silicon Macs. API customers can use Zero Data Retention. Kimi Enterprise administrators can switch the model on for their organizations, but access starts off by default.
Model IDs for developers
According to Unite.AI, the API lists the models as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Inside Kimi Code, users pick a model with the /model command. The Kimi API platform model list is the place to confirm current identifiers before you wire anything into production.
# Kimi Code: open the model picker
/model
# Model IDs reported for the API
kimi-k3.1-pro
kimi-k3.1-pro-highspeedSpecs: Context Window, Parameters, and Inputs
The 1-million-token context window is the one spec Moonshot confirms in its own words. Everything else on the hardware side comes from reporting. The Decoder says K3.1 Pro keeps the same dimensions as K3, with 2.8 trillion total parameters and 104 billion active ones. That fits the 3T-class label, and the gap between 2.8 trillion total and 104 billion active parameters implies a sparse design where only a slice of the network runs per token.
The Decoder also reports that the model handles text and images. OpenRouter calls it multimodal. Treat the parameter counts as reported figures until Moonshot publishes a model card alongside the weights.
Unite.AI notes that Kimi Code documentation lists model IDs across the K3.1 Pro, K3, K2.8 Preview and K2.7 Code Highspeed families. It describes k3.1-pro as a 3T-parameter flagship with a context window of up to 1,048,576 tokens for higher-tier members. That is 2^20 tokens, slightly above one million, which is worth remembering when you size prompts.
Benchmark Results: Where K3.1 Pro Leads and Trails
Moonshot published its own benchmark table, and The Decoder reproduced it. Every result uses reasoning effort set to max. These are company-reported numbers, so independent replication matters. The table below pulls the rows that matter most for coding, research and agent work.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | DeepSeek-V4.1-Flash |
|---|---|---|---|---|
| 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 |
| HLE-Full (with tools) | 70.0 | 67.7 | 57.2 | 58.8 |
| BrowseComp | 92.1 | 90.8 | 91.5 | n/a |
| OfficeQA Pro v2 | 64.4 | 62.2 | 54.4 | 40.4 |
| GDPval-AA v2.1 (Elo) | 1,846 | 1,866 | 1,573 | 1,746 |
| GPQA-Diamond | 94.2 | 93.9 | 96.1 | 92.2 |
| ExploitBench | 74.7 | 91.0 | 100 | n/a |
Source: Moonshot’s benchmark table as reproduced by The Decoder. The full table also lists Claude Fable 5.1 and GLM-5.3.
Read the table as a split decision. By my count, K3.1 Pro posts the top score on four of the ten rows above: SWE Marathon, HLE-Full, BrowseComp and OfficeQA Pro v2. It beats Opus 5.5 on GPQA-Diamond too, but GPT-6 Astra takes that row at 96.1. K3.1 Pro trails Opus 5.5 on Terminal-Bench by 0.7 points, on BenchCAD by 3.5, and on GDPval-AA by 20 Elo. DeepSeek-V4.1-Flash edges it on DeepSWE.
The widest gap sits at the bottom. ExploitBench shows K3.1 Pro at 74.7 against 91.0 for Opus 5.5. That gap is deliberate, and the cybersecurity section below explains why.
The Artificial Analysis Index: A Fourth-Place Debut
Independent scoring tells a cleaner story than vendor tables. Artificial Analysis, cited by The Decoder, ranks K3.1 Pro fourth with 55 points. If Moonshot ships the weights as planned, The Decoder says K3.1 Pro would become the highest-scoring open-weights model available, and the gap between open and proprietary models would shrink from 12 points to 3.
| Rank | Model | Intelligence Index score | Note |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 58 | Top proprietary model |
| 2 (tie) | Claude Sonnet 5.5 | 56 | Proprietary |
| 2 (tie) | Muse Wonder Full | 56 | Meta’s frontier model |
| 4 | Kimi K3.1 Pro | 55 | Weights release planned |
| Lower | GLM-5.3 | 45 | Cited by The Decoder as a reference point |
| Lower | Kimi K3 | 44 | Predecessor (55 minus the 11-point gain) |
Source: Artificial Analysis figures as reported by The Decoder. The Kimi K3 score is derived from the reported 11-point jump.
The Decoder adds that no model of comparable parameter count scores higher on the Index, which puts K3.1 Pro on the Pareto line for intelligence per dollar. For a related snapshot of how the top proprietary models stack up, see our LLM leaderboard comparing Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon.
Pricing: A 33% Cut That Barely Moves Task Cost
Moonshot’s own materials do not state a price in the sources we could check, so every number here comes from The Decoder. It reports that token prices drop by 33 percent against K3. Input falls from $3 to $2 per million tokens, output from $15 to $10, and cache hits from $0.30 to $0.20.
| Item | Kimi K3 | Kimi K3.1 Pro | Kimi K3.1 Pro Highspeed |
|---|---|---|---|
| Input, per million tokens | $3.00 | $2.00 | $4.00 |
| Output, per million tokens | $15.00 | $10.00 | $20.00 |
| Cached input, per million tokens | $0.30 | $0.20 | $0.20 |
| Speed claim | Baseline | Baseline | 2x to 3x faster |
| Cost per Index task | $2.00 | $1.98 | Not reported |
Source: The Decoder, citing Moonshot pricing and Artificial Analysis task costs. Verify current rates on the official API platform before budgeting.
The last row is the interesting one. Despite the headline cut, an Artificial Analysis task costs $1.98 on K3.1 Pro versus $2.00 on K3. The reason, per The Decoder, is that the new model generates 14 percent more output tokens. A lower sticker price does not guarantee a lower bill. Reasoning-heavy models burn more tokens, so teams should measure cost per finished task, not cost per token. Our comparison of Gemini, ChatGPT and Claude API cache pricing shows how much the cached-input rate alone can change a long-context workload.
The Highspeed tier doubles the per-token price in exchange for the speed claim. That is a classic latency premium, and it suits interactive coding agents better than overnight batch jobs.
Agentic and Long-Horizon Coding Gains
Agentic work shows the biggest improvement. The Decoder reports that K3.1 Pro reaches 1,846 Elo on GDPval-AA, 312 points above K3. On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra at 58.2 and just behind Claude Opus 5.5 at 66.4. K3 managed only 13 percent on the same test.
Unite.AI adds an internal benchmark. On Kimi Code Bench 3.0 Long-horizon, which covers multi-step execution, tool use and end-to-end problem solving, K3.1 Pro scored 61.2 percent against 55.95 percent for K3. Because Moonshot built that test, outsiders cannot yet check it.
The most vivid example in the blog post is a physics project. According to Unite.AI, K3.1 Pro built a human walking dynamics simulator on top of an open-source anatomical model with 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, it refined a walking controller. When reward shaping failed to fix arm motion, it probed the simulator’s dynamics directly, found 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller. That is one curated demo, not a reliability statistic, but it shows the long-horizon behavior Moonshot is selling.
Hallucination numbers moved too. On Moonshot’s internal Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4 percent with K3 to 8.3 percent. The company stresses that the test is built to provoke errors and does not measure everyday error rates. The Decoder notes GPT-6 Astra still scores 5.5 percent on the same benchmark.
3D Creation and Financial Research
Two of the four headline capabilities fall outside classic chatbot territory. On 3D creation, The Decoder says K3.1 Pro works directly with Unity, Unreal Engine and Godot, handling both quick fixes and full project builds. It can write and run scripts for Blender and CAD software without MCP connectors. Moonshot also launched a “3D Make” feature on Kimi.com that lets users preview, edit and export models, and Unite.AI says downloads come ready for 3D printing.
On BenchCAD without tools, the model hits 72.8 percent, nearly level with Claude Opus 5.5 at 73.0. With tools it reaches 92.7 percent, which still trails Opus 5.5 and GPT-6 Astra by a few points.
On finance, Moonshot reports gains over K3 of 7.2 points on Financial Modeling and 11.6 points on Financial Research. Unite.AI adds that Business Operation performance rose to about 4.2 times K3’s level. The model also posts 64.4 on OfficeQA Pro v2, a test of fiscal data analysis over real financial documents, which tops every model in the comparison table. These are internal benchmarks, so finance teams should run their own spreadsheets through the model before trusting the headline.
Cybersecurity Controls Behind the Open-Weight Release
Moonshot devoted an entire section of its technical blog to cybersecurity, and that section explains the ExploitBench gap. The company says it saw vulnerability-research ability keep rising as post-training scaled. It traced the trend to cyber-related content in the general mid-training corpus and to related tasks in the reinforcement learning mix.
What the data removal changed
For the open-weight release, Moonshot removed cyber-related data from both training stages while keeping the general coding tasks the model needs. On its in-house xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software such as the Linux kernel, Node.js and JavaScriptCore, a checkpoint trained without the safeguards solves 90.5 percent of problems at pass@3. The final K3.1 Pro solves 81.0 percent. The Decoder lists DeepSeek-V4.1-Flash at 85.7 percent on the same test and K3 at 33.3 percent.
Moonshot states two principles. It does not intend to release open weights that extend the absolute frontier of cyber capability. It also wants to raise open-weight cyber capability gradually, because wider access helps more teams audit code they maintain, including legacy systems that rarely get dedicated reviews. The company says K3.1 Pro’s cyber ability stays behind proprietary frontier models that have been public for at least three months.
Skeptics will note the obvious limit. Removing data reduces capability, but a determined party could fine-tune it back in after the weights go public. Moonshot has not claimed otherwise, and no independent test of that scenario exists yet.
Open Weights and the November 4 Question
The “open-weight” label carries a caveat today. The model runs through hosted products now, but the full weights have not been published. TokenPost reported that Moonshot plans a weight release before November 4, and The Tribune carried a similar account of a plan to release full weights by November. The Decoder gives the same November 4 date.
We could not confirm that date in the official Kimi materials we reviewed, so treat it as a reported target. The distinction matters for planning. A team that wants to self-host a 2.8-trillion-parameter model needs serious GPU capacity, and the license terms will decide whether commercial use is allowed. Moonshot has not detailed either in the sources available to us.
Until the weights land, K3.1 Pro behaves like any hosted frontier API. After they land, hosting providers can compete on price and latency, which is where open weights change the market.
Competitive Comparison: Opus 5.5, GPT-6 Astra and DeepSeek
Against Claude Opus 5.5, K3.1 Pro closes most of the gap on agentic and research tasks while giving up ground on cyber and CAD work. The three-point Index deficit is small. Opus 5.5 also leads on GDPval-AA, Terminal-Bench and BenchCAD in Moonshot’s own table, so Moonshot is not claiming a clean sweep.
Against GPT-6 Astra, the picture splits. Kimi leads on Terminal-Bench, SWE Marathon, HLE-Full and OfficeQA Pro v2. GPT-6 Astra wins GPQA-Diamond, BenchCAD and the hallucination stress test. On GDPval-AA the Elo gap is large, 1,846 for Kimi versus 1,573.
The closest rival by philosophy is DeepSeek-V4.1-Flash. It beats K3.1 Pro on DeepSWE (77.0 versus 75.5) but trails on nearly everything else in the table. DeepSeek also has fresh capital behind it, as we covered in our report on DeepSeek’s reported $15 billion raise ahead of an IPO. Two well-funded Chinese labs now push open-weight quality toward the closed frontier from different directions.
Historical Context: From K2.7 to K3.1
Moonshot ships several variants side by side. Unite.AI’s reading of the Kimi Code docs shows K2.7 Code Highspeed, K2.8 Preview, K3 and now K3.1 Pro all sitting in one model menu. The company ships point releases and speed variants instead of waiting for one giant generational jump.
The K3.1 approach differs from a fresh pretraining run. By keeping K3’s architecture and spending more compute on post-training and reinforcement learning, Moonshot lifted the Index score by 11 points without changing parameter counts. The blog credits upgrades to the reinforcement learning framework, long-context inference, sandbox infrastructure and rollout scheduling. If that account holds, the lesson for the industry is that post-training scale, not just model size, now drives the leaderboard.
Market Impact for Developers and Enterprises
The immediate winners are teams that already route traffic across providers. With K3.1 Pro on OpenRouter from day one, switching costs are low and price pressure on closed models rises. The Decoder’s own top-stories list on launch day pointed to a related trend, which we examined in our look at MiniMax M3.1 Flash and its accuracy questions. Chinese labs keep cutting prices while narrowing quality gaps.
Enterprises face a different calculation. Kimi Enterprise ships with the model disabled by default, and Zero Data Retention on the API addresses one common compliance objection. Many regulated buyers will still hesitate over data residency and vendor origin. Those concerns are policy questions, not benchmark questions, and no score changes them.
For anyone tracking the wider field, our AI and machine learning coverage hub collects the model launches that shaped the last quarter.
Five Predictions for the K3.1 Series
These are analyst judgments, not reported facts.
- More K3.1 models will follow. Moonshot called Pro the first model in the series, which signals smaller or specialized siblings.
- Hosted prices for K3.1 Pro will fall after the weights ship. Third-party inference providers typically undercut the originator once weights are public.
- Independent cyber evaluations will test the data-removal claim. Security researchers will probe whether fine-tuning restores the stripped capability.
- Latency tiers will spread. Highspeed variants at roughly double the token price give rivals a template to copy.
- Enterprise pilots will concentrate on finance and coding. Those are the two areas where Moonshot published its largest internal gains.
FAQ
What is Kimi K3.1 Pro? It is Moonshot AI’s newest model and the first in the K3.1 series. Moonshot calls it its most capable 3T-class open-weight model to date, with a 1-million-token context window.
When was it released? Official Kimi materials carry an October 23, 2026 publication date.
What is Kimi K3.1 Pro Highspeed? It is a faster version launched alongside the standard model. Moonshot says it runs two to three times faster with the same capabilities, according to The Decoder and Unite.AI.
Where can I use it? Both models are available on Kimi.com, Kimi Work, Kimi Code and the Kimi API. OpenRouter also lists the standard model.
How much does it cost? Moonshot’s materials we reviewed give no price. The Decoder reports $2 per million input tokens and $10 per million output tokens for K3.1 Pro, and $4 and $20 for Highspeed. Confirm on the API platform.
Are the model weights available now? Not yet. Outlets including TokenPost and The Decoder report a planned release before or by November 4, but the official Kimi materials we reviewed do not confirm the date.
How does it rank against Claude Opus 5.5? The Decoder reports 55 points on the Artificial Analysis Intelligence Index against 58 for Claude Opus 5.5, placing Kimi K3.1 Pro fourth overall.
