The open-weight race just got a lot closer. On October 23, 2026, Chinese AI lab Moonshot AI released Kimi K3.1 Pro, the first model in its new K3.1 series — and the first independent scoreboard put it fourth in the world, within three points of the best proprietary model on the planet.
Moonshot calls it "our most capable 3T-class open-weight model to date," and the headline spec is a 1-million-token context window — enough to hold an entire codebase, a shelf of financial filings, or a long research project in a single conversation. A faster sibling, Kimi K3.1 Pro Highspeed, launched the same day. The full model weights are reported to follow by November 4, which would make K3.1 Pro the highest-scoring open-weights model ever released.
This guide covers what Moonshot actually announced, what only outside outlets have reported, where K3.1 Pro beats Claude Opus 5.5 and GPT-6 Astra on Moonshot's own benchmark table (and where it clearly loses), what it costs when you run the arithmetic honestly, how to try it today, and who should — and should not — build on it yet.
Kimi K3.1 Pro at a glance
| Item | Detail |
|---|---|
| Developer | Moonshot AI (Beijing, China) |
| Released | October 23, 2026 |
| Series position | First model in the new K3.1 series |
| Model class | 3T-class open-weight model (reported: 2.8 trillion total parameters, 104 billion active) |
| Context window | 1 million tokens (Moonshot-confirmed; up to 1,048,576 tokens for higher-tier members, per Kimi Code docs) |
| Inputs | Text and images (multimodal, per OpenRouter listing) |
| Sibling model | Kimi K3.1 Pro Highspeed — same capabilities, 2–3x faster (reported) |
| Where to use it | Kimi.com, Kimi Work, Kimi Code, Kimi API, and OpenRouter |
| API model IDs (reported) |
kimi-k3.1-pro, kimi-k3.1-pro-highspeed
|
| Independent score | 55 on the Artificial Analysis Intelligence Index — 4th overall |
| Price (reported) | $2 per million input tokens, $10 per million output tokens, $0.20 cached input |
| Open weights | Not yet published; reported target before/by November 4, 2026 (unconfirmed by Moonshot's own materials) |
| Built for | Long-horizon coding, agentic work, 3D creation, financial research and modeling |
What Moonshot AI actually announced
It is worth separating what Moonshot said from what everyone else is saying, because this launch has both in unusual abundance. The official announcement, dated October 23, 2026, confirms four things: K3.1 Pro is the first model in the K3.1 series; it is a 3T-class open-weight model; it carries a 1-million-token context window; and it targets coding, 3D creation, financial research, and long-horizon agent work. Moonshot's own summary of the improvement is direct: "It delivers substantial improvements across coding, 3D creation, financial research, and other agentic capabilities."
Everything about parameter counts, pricing, and the weights date comes from reporting by outlets such as The Decoder, Unite.AI, and TokenPost — not from Moonshot's own announcement materials. That distinction matters, and this guide keeps it visible throughout: where a number is reported rather than confirmed, it says so.
The lineage is also important. K3.1 Pro is not a from-scratch pretraining run. Multiple outlets describe it as a post-trained version of Kimi K3, built with substantially larger post-training compute. OpenRouter, which listed the model on launch day, describes it in nearly identical terms: a 1M-context, multimodal open-weight model, post-trained from Kimi K3 for long-horizon coding and agentic work. The result of that post-training push is the story of this launch: the same architecture, 11 points higher on the independent index. If that account holds, it is evidence for something the whole industry has been debating — that post-training scale, not just raw model size, is now driving the leaderboard.
The 1-million-token context window, explained
A context window is how much text and data a model can hold in working memory at once. One million tokens is roughly 750,000 English words — about ten average novels, or a medium-sized software repository with its documentation, or several years of a company's financial filings. Most flagship chatbots still work with far less in their default settings, so a 1M window changes what kinds of jobs you can hand to a model in one go.
For developers, the practical meaning is whole-codebase reasoning: you can point Kimi Code at a repository and ask it to trace a bug across dozens of files, refactor a module with full awareness of its callers, or review a pull request against the entire project history. For researchers and analysts, it means dropping a stack of reports, transcripts, or filings into one session and asking questions that require connecting facts across all of them. Kimi Code documentation reportedly lists the ceiling at 1,048,576 tokens — exactly 2 to the power of 20 — for higher-tier members, which is worth remembering when you size prompts: the "1 million" headline is a rounded figure, and the precise limit depends on your tier.
There is a catch that applies to every long-context model, and K3.1 Pro is no exception. A big window does not guarantee the model actually uses all of it well. Long-context recall — finding the one needle fact buried in a haystack of documents — is a separate skill from merely accepting a long input, and independent long-context benchmarks for K3.1 Pro have not yet landed. Treat the 1M figure as capacity, not as a promise of perfect recall, until third-party tests confirm it.
Kimi K3.1 Pro Highspeed: the same brain, faster
The second product in the launch is easy to misunderstand, so Moonshot's framing deserves repeating: Highspeed is a latency tier, not a smaller or weaker model. Per Unite.AI and The Decoder, Moonshot says it runs two to three times faster than standard K3.1 Pro with the same capabilities. The price of that speed is steep — reportedly double the per-token rate — which makes Highspeed a tool for interactive work (a coding agent you are watching type in real time, a customer-facing assistant that cannot keep users waiting) rather than overnight batch jobs, where standard K3.1 Pro's lower price wins.
This two-tier pattern is one to watch across the industry. If latency tiers at roughly double the price become standard, the "which model" question splits into two questions: which model, and how fast do you need it?
Where both models run today
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." The detail around that statement, reported by Unite.AI, fills in the picture:
- Kimi.com — the web chatbot, where a new "3D Make" feature lets users preview, edit, and export 3D models.
- Kimi Work — the desktop app (version 3.1.0 or later on Windows and Apple silicon Macs).
-
Kimi Code — the agentic coding tool, where users pick a model with the
/modelcommand. -
Kimi API — with Zero Data Retention available for API customers, and model IDs reported as
kimi-k3.1-proandkimi-k3.1-pro-highspeed. - OpenRouter — listed the standard model on launch day, which matters because it lets teams route traffic across providers without rewriting integrations.
- Mobile — iOS, Android, and HarmonyOS apps.
One enterprise detail stands out: Kimi Enterprise administrators can switch the model on for their organizations, but access starts off by default. That is a deliberate choice — enterprises opt in rather than waking up to a new model in their stack — and it tells you Moonshot is pitching this launch at careful corporate buyers, not just enthusiasts.
Benchmarks: where K3.1 Pro leads and where it trails
Moonshot published its own benchmark table, reproduced by The Decoder, with every result measured at maximum reasoning effort. These are company-reported numbers, so independent replication matters — but the table is unusually honest, because it shows K3.1 Pro losing several rows. Read it as a split decision, not a sweep.
| 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.
Count the wins: K3.1 Pro posts the top score on four of the ten rows — SWE Marathon, HLE-Full, BrowseComp, and OfficeQA Pro v2 — and edges Opus 5.5 on GPQA-Diamond, though GPT-6 Astra takes that row outright. It trails Opus 5.5 narrowly 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 and a perfect 100 for GPT-6 Astra. That gap is deliberate, and the cybersecurity section below explains why.
Two rows deserve special attention because they measure the things Moonshot is actually selling. On Terminal-Bench 4.0, a test of real terminal-based coding tasks, K3.1 Pro scores 65.7 percent — ahead of GPT-6 Astra's 58.2 and just behind Opus 5.5's 66.4. Its predecessor K3 managed only 13 percent on the same test, which gives you a sense of how large the post-training jump really is. On GDPval-AA, an agentic evaluation, K3.1 Pro reaches 1,846 Elo — 312 points above K3, and far ahead of GPT-6 Astra's 1,573. For teams buying an agent rather than a chatbot, those two rows are the ones that matter.
The independent score: fourth in the world
Vendor tables are useful, but independent scoring tells a cleaner story. Artificial Analysis, cited by The Decoder, ranks K3.1 Pro fourth overall with 55 points on its Intelligence Index — an 11-point jump over Kimi K3's 44. Only three models sit above it: Claude Opus 5.5 at 58, and Claude Sonnet 5.5 and Meta's Muse Wonder Full tied at 56.
| 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 number that matters most is not 55 but 3. The Decoder notes that if Moonshot ships the weights as planned, 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. Three points is within the margin where a team's choice stops being about raw intelligence and starts being about price, latency, data control, and self-hosting — which is exactly the territory where open weights win. The Decoder adds that no model of comparable parameter count scores higher on the Index, putting K3.1 Pro on the Pareto line for intelligence per dollar.
There are caveats, and Artificial Analysis itself supplies one: its analysts found K3.1 Pro slower than average and very wordy. Verbosity is not a cosmetic flaw — it directly inflates cost, because output tokens are the expensive ones. That connects to the pricing section, where the headline cut turns out to be smaller than it looks.
Pricing: a 33% cut that barely moves the bill
Moonshot's own announcement materials do not state a price in the sources reviewed for this guide, so every number here comes from reporting — The Decoder, citing Moonshot pricing, and the Kimi API platform's pricing documentation. Treat the table as reported, and verify current rates on the official API platform before budgeting.
| 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 Artificial Analysis task | $2.00 | $1.98 | Not reported |
Source: The Decoder, citing Moonshot pricing and Artificial Analysis task costs.
The headline is a 33 percent cut on every rate. The reality, in the last row, is more interesting: an Artificial Analysis task costs $1.98 on K3.1 Pro versus $2.00 on K3 — essentially unchanged. 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 when the model talks more. Reasoning-heavy models burn more tokens, so teams should measure cost per finished task, not cost per token — and should test with their own prompts before migrating a workload, because verbosity varies by task type.
A worked example makes the point concrete. Suppose your coding agent processes 10 million input tokens and generates 2 million output tokens in a month, with half the input hitting the cache. On the reported K3.1 Pro rates, the arithmetic is: 5 million fresh input tokens at $2.00 per million = $10.00; 5 million cached input tokens at $0.20 per million = $1.00; 2 million output tokens at $10.00 per million = $20.00; total $31.00. The same workload on the old K3 rates would be $15.00 + $1.50 + $30.00 = $46.50. That is a genuine saving of about a third — but only if output volume stays flat. If K3.1 Pro's verbosity adds 14 percent more output tokens, output rises to 2.28 million tokens and $22.80, pushing the total to $33.80. Still cheaper, but the saving shrinks from 33 percent to about 27 percent. And if your workload is output-heavy — long code generation rather than long-context reading — the saving shrinks further. Cache discipline matters enormously here: cached input at $0.20 is fifty times cheaper than output at $10.00, so keeping repeated context (system prompts, repository context, reference docs) in cache is the single biggest lever on the bill.
Highspeed doubles the per-token price for its speed claim. That is a classic latency premium, and it suits interactive coding agents better than overnight batch jobs. For a batch pipeline that runs while nobody watches, standard K3.1 Pro at half the price is the rational choice; for a pair-programming session where every second of waiting breaks flow, Highspeed may pay for itself in developer time.
For context, our best AI coding assistants guide compares the tools this model will now power, and our DeepSeek-V4.1-Flash guide covers the closest rival by philosophy — another open-leaning lab cutting prices while chasing the frontier.
Agentic and long-horizon coding: the biggest jump
If one theme runs through this launch, it is long-horizon work — tasks that take many steps, use tools, recover from errors, and finish the job without a human steering every move. Moonshot's own internal benchmark, Kimi Code Bench 3.0 Long-horizon, which covers multi-step execution, tool use, and end-to-end problem solving, puts K3.1 Pro at 61.2 percent against 55.95 percent for K3. Because Moonshot built that test, outsiders cannot yet check it — but the direction matches the independent rows: SWE Marathon, Terminal-Bench, and GDPval-AA all moved sharply in the same direction.
The most vivid example in Moonshot's blog post is a physics project, reported by 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 — and 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, and it should be read as such. But it shows the behavior Moonshot is selling: not answering a question, but pursuing a goal across hundreds of steps, noticing its own approach is failing, diagnosing why, and changing course. That is the difference between a chatbot and an agent, and it is where K3.1 Pro's gains are concentrated.
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 — a fair warning, since stress-test numbers always look worse than daily use. The Decoder notes GPT-6 Astra still scores 5.5 percent on the same benchmark, so K3.1 Pro has narrowed but not closed that particular gap.
For developers weighing this against the current coding-model field, the honest summary is: K3.1 Pro is now a credible primary coding model, not just a cheap fallback. Its predecessor was not. The Terminal-Bench jump from 13 percent to 65.7 percent is the single most important number in the launch, because it measures real terminal work rather than multiple-choice knowledge. Teams already using Claude Duet or other agentic setups should still test before switching — agent frameworks, tool integrations, and prompt habits all interact with the model — but K3.1 Pro has earned a place in the evaluation, which K3 had not.
3D creation and financial research: beyond the chatbot
Two of the four headline capabilities fall outside classic chatbot territory, and they are the ones that make this launch interesting to people who would never buy a "coding model."
On 3D creation, The Decoder reports that 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 — meaning the model drives the tools itself rather than needing a plugin bridge. Moonshot also launched a "3D Make" feature on Kimi.com that lets users preview, edit, and export models, and Unite.AI reports that 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. The gap between tool-assisted and unassisted scores is itself informative: this model gets dramatically better when it can actually run the software, which is the agentic story again.
On finance, Moonshot reports gains over K3 of 7.2 points on Financial Modeling and 11.6 points on Financial Research, and 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 — financial work punishes confident errors more than most domains, and a benchmark win is not a compliance argument. But combined with the 1M context window, the pitch is clear: an analyst can load years of filings, earnings calls, and market data into one session and ask questions that span all of them.
The cybersecurity decision behind the open-weight release
Moonshot devoted an entire section of its technical blog to cybersecurity, and that section explains the ExploitBench gap — the one row where K3.1 Pro trails badly. The company says it saw vulnerability-research ability keep rising as post-training scaled, traced the trend to cyber-related content in the general mid-training corpus and to related tasks in the reinforcement learning mix, and made a deliberate choice about the open-weight release.
For the open-weight version, Moonshot removed cyber-related data from both training stages while keeping the general coding tasks the model needs. The effect, on its in-house xdev-bench-hard — which reproduces recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, and JavaScriptCore — is measurable: a checkpoint trained without the safeguards solves 90.5 percent of problems at pass@3, while 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, so K3.1 Pro still sits far above its predecessor even after the removal.
Moonshot states two principles behind the decision. It does not intend to release open weights that extend the absolute frontier of cyber capability. And it 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, and it is worth stating plainly: 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. The safeguard raises the bar; it does not make the capability unrecoverable. For security teams, the practical takeaway is that K3.1 Pro is a strong defensive tool — auditing your own code, finding vulnerabilities in systems you own — while its offensive ceiling has been deliberately kept below the closed frontier. Whether that balance holds after the weights ship is one of the most important open questions of this launch, and independent cyber evaluations will test it within weeks of November 4.
Open weights and the November 4 question
The "open-weight" label carries a caveat today, and buyers should understand it before planning around it. The model runs through hosted products now — Kimi.com, Kimi Work, Kimi Code, the API — 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. Moonshot's own announcement materials reviewed for this guide do not confirm that date, so treat it as a reported target, not a promise.
The distinction matters for planning in two ways. First, a team that wants to self-host a 2.8-trillion-parameter model needs serious GPU capacity — this is not a model you run on a laptop, and the infrastructure bill may exceed the API bill for all but the largest deployments. Second, the license terms will decide whether commercial use is allowed, whether fine-tuning and redistribution are permitted, and what attribution is required. Moonshot has not detailed either the hardware guidance or the license in the sources available to us. Until the weights land, K3.1 Pro behaves like any hosted frontier API: convenient, priced per token, and governed by the provider's terms. After they land, hosting providers can compete on price and latency, which is where open weights change the market — the same dynamic that made DeepSeek's releases so disruptive.
If you are deciding between starting on the API now versus waiting for weights, the pragmatic answer is to start on the API now and design your integration so the model is swappable. K3.1 Pro is already on OpenRouter, which makes that swap nearly free. Waiting three weeks to save nothing, when the API is live today, is rarely the right call — unless your use case legally requires self-hosting, in which case the wait is mandatory and the November 4 date is your planning anchor.
Kimi K3.1 Pro vs the field
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, and Opus 5.5's leads on GDPval-AA, Terminal-Bench, and BenchCAD are narrow — but they are real, and they sit in exactly the areas where Opus has always been strongest. For teams already deep in the Claude ecosystem, K3.1 Pro is a credible second provider for routing and cost control rather than an automatic replacement. Our ChatGPT vs Claude vs Gemini comparison covers how these flagship families differ in everyday use.
Against GPT-6 Astra, the picture splits by task type. 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 — which suggests Kimi is the stronger agent while Astra remains the stronger general reasoner on some academic-style tests. The choice between them depends on whether your workload looks more like an agent's job or an exam.
Against DeepSeek-V4.1-Flash, the closest rival by philosophy, K3.1 Pro wins nearly everywhere in Moonshot's table except DeepSWE (77.0 versus 75.5). DeepSeek also has fresh capital behind it, with a 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 — DeepSeek on efficiency and price, Moonshot on agent capability and context length. That competition is good news for everyone buying models, because it keeps pressure on both quality and price.
Against Google's Gemini 4 Argon, the other October frontier release, the comparison is less direct: Gemini's headline strengths are agent benchmarks of a different kind and deep Google ecosystem integration, covered in our Gemini 4 Argon guide. K3.1 Pro's edge is openness — the weights path — while Gemini's edge is integration. Teams choosing between them are really choosing between an open-weights future and a platform future.
Who should use Kimi K3.1 Pro now — and who should wait
Use it now if you are:
- A developer or team running agentic coding workflows, especially long multi-step tasks where K3's weakness was fatal and K3.1 Pro's Terminal-Bench jump changes the calculus. Try it in Kimi Code alongside your current model and compare on your own repository.
- A team with very long documents — codebases, filing sets, research corpora — where the 1M context window removes a real bottleneck today.
- A cost-sensitive team routing across providers via OpenRouter, where K3.1 Pro's reported $2/$10 pricing and 55-point index score give you near-frontier quality at a mid-tier price.
- A 3D or CAD workflow user curious about model-driven Blender or engine scripting, where the tool-use capability is the feature, not the chat.
Wait or proceed carefully if you are:
- A regulated enterprise that requires self-hosting or specific data-residency guarantees. The weights are not out yet, the license is unknown, and Kimi Enterprise starts with the model disabled by default for a reason.
- A security team evaluating offensive cyber capability. Moonshot deliberately capped this in the open-weight version, and the fine-tuning question is unresolved.
- Anyone whose workload is output-heavy and latency-insensitive. Test your real prompts first: verbosity may eat the headline price cut.
- Anyone who needs guaranteed long-context recall at the full 1M window. Independent long-context benchmarks have not landed yet.
How to try Kimi K3.1 Pro today
- Fastest path: open Kimi.com and select Kimi K3.1 Pro in the model picker. No installation, and the 3D Make feature is available there.
-
For coding: install Kimi Code, open the model picker with the
/modelcommand, and choose K3.1 Pro (or Highspeed for interactive sessions). Point it at a real repository, not a toy example — the long-horizon gains only show on real work. -
For API use: check the model list on the Kimi API platform and confirm the current rates and model IDs (
kimi-k3.1-pro,kimi-k3.1-pro-highspeed) before wiring anything into production. Zero Data Retention is available for API customers. - For multi-provider routing: find the model on OpenRouter and compare its live price and latency against your current provider on identical prompts.
- For enterprises: ask your Kimi Enterprise administrator to enable the model for a pilot group first, and measure cost per finished task — not cost per token — for two weeks before rolling out wider.
- For the weights: mark November 4 on your calendar as a reported target, not a confirmed date, and do not commit self-hosting budgets until the license terms are public.
What is still not confirmed
Honest coverage requires listing what we do not know yet, because this launch has more reported-than-confirmed facts than most:
- The November 4 weights date — reported by TokenPost, The Tribune, and The Decoder; not confirmed in Moonshot's own announcement materials.
- Pricing — reported by The Decoder and the API platform documentation; Moonshot's announcement materials reviewed here state no price.
- Parameter counts — 2.8 trillion total and 104 billion active, reported by The Decoder; consistent with the "3T-class" label but not confirmed in Moonshot's own words.
- Long-context recall quality at the full 1M window — no independent benchmark published yet.
- License terms for the weights — unknown until publication.
- Whether fine-tuning restores the removed cyber capability — untested, and the key safety question after the weights ship.
None of these gaps makes the launch less significant. They define the work of the next month: independent benchmarks, the weights release, license publication, and security probing will convert today's reported facts into confirmed ones — or correct them.
What happens next: five things to watch
These are analyst judgments, not reported facts — included so you can check this guide against events rather than take predictions as news.
- More K3.1 models will follow. Moonshot called Pro the first model in the series, which signals smaller or specialized siblings — a faster tier already exists, and a cheaper or coding-specialized variant would fit the pattern.
- Hosted prices for K3.1 Pro will fall after the weights ship. Third-party inference providers typically undercut the originator once weights are public, and OpenRouter already makes that competition visible.
- Independent cyber evaluations will test the data-removal claim. Security researchers will probe whether fine-tuning restores the stripped capability, and their findings will shape how enterprises treat open-weight releases.
- Latency tiers will spread. Highspeed variants at roughly double the token price give rivals a template to copy, and expect similar tiers from other labs within a quarter.
- Enterprise pilots will concentrate on finance and coding. Those are the two areas where Moonshot published its largest internal gains, and where the 1M context window has the clearest business case.
The wider context is worth noting too. October 2026 has been an unusually crowded model month — MiniMax M3.1 Flash, Nemotron 3.5 Bolt, Gemini 4 Argon, Claude Haiku 5.5, and now Kimi K3.1 Pro, covered in our Nemotron 3.5 Bolt guide. The pattern across all of them is the same: prices falling, context windows growing, and agent capability becoming the battleground. K3.1 Pro is the strongest evidence yet that open weights are not a compromise tier anymore — they are three points from the top of the world.
Frequently asked questions
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, built for coding, 3D creation, financial research, and long-horizon agent work.
When was Kimi K3.1 Pro released?
October 23, 2026. Moonshot's official announcement materials carry that date, and the model was available the same day on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
What is Kimi K3.1 Pro Highspeed?
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, at reportedly double the per-token price.
Where can I use Kimi K3.1 Pro?
On Kimi.com, Kimi Work, Kimi Code, and the Kimi API, plus OpenRouter. Mobile apps cover iOS, Android, and HarmonyOS.
How much does Kimi K3.1 Pro cost?
Moonshot's announcement materials reviewed here 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, with cached input at $0.20. Confirm current rates on the Kimi API pricing page before budgeting.
Are the model weights available now?
Not yet. Outlets including TokenPost and The Decoder report a planned release before or by November 4, 2026, but Moonshot's own announcement materials do not confirm the date.
How does Kimi K3.1 Pro rank against Claude Opus 5.5?
Artificial Analysis, as reported by The Decoder, scores K3.1 Pro at 55 on its Intelligence Index against 58 for Claude Opus 5.5, placing Kimi fourth overall — and, if the weights ship as reported, the highest-scoring open-weights model available.
Is Kimi K3.1 Pro really open source?
Careful with the word. Moonshot calls it an open-weight model, which means the weights are planned for publication — not necessarily that the training code, data, and full license are open. Until the license terms are published, "open-weight" is the accurate label, and "open source" is a claim to verify.
Should I switch my coding workflow to Kimi K3.1 Pro?
Test it on your own repository first. The Terminal-Bench and agentic gains are large, but agent frameworks, tool integrations, and your own prompt habits all interact with the model. Run it alongside your current model for two weeks and compare cost per finished task, not just benchmark scores.
What is Moonshot AI?
A Beijing-based AI lab behind the Kimi chatbot and the Kimi model family. Its K2 series gained attention earlier in 2026 for agentic coding, and the K3 series — now K3.1 — is its push toward the frontier, with open weights as its distinguishing strategy against closed American labs.
Sources and further reading
- Moonshot AI, Kimi K3.1 Pro announcement, October 23, 2026 — kimi.ai
- Kimi API platform, model pricing documentation — platform.kimi.ai
- Artificial Analysis, Kimi K3.1 Pro model page — artificialanalysis.ai
- CGTN, "China's new AI model advances AI race," October 24, 2026 — news.cgtn.com
- The Decoder and Unite.AI launch coverage, October 24, 2026, via Tech Insider's Kimi K3.1 Pro analysis
This guide distinguishes Moonshot-confirmed facts from reported figures throughout. Prices, parameter counts, and the November 4 weights date are reported by outside outlets and should be re-verified on the official Kimi platform before production or budget decisions.
Top comments (0)