Update, October 23, 2026, 17:15 UTC: Kimi K3.1 is out. Moonshot released Kimi K3.1 Pro, “our most capable 3T-class open-weight model to date”, and announced it on X. It costs $2.00 per million input tokens and $10.00 output on Moonshot’s API, a third less than K3, and in Moonshot’s own runs it scores 65.7 on Terminal-Bench 4.0 against Claude Opus 5.5’s 66.4. It is live today in Kimi, Kimi Work, Kimi Code and the API, and open weights are promised by November 4. The release record and a scorecard of every earlier claim are below.
Update, October 23, 2026, 01:41 UTC: Moonshot put the name on screen itself. At 01:41 UTC Kimi’s X account posted a video with no caption: Kimi Code on an old CRT monitor, the session’s model set to Kimi K3.1 with thinking on high, and a prompt asking it to build a complex procedural 3D scene in Three.js. The same frame shows Kimi Code version 2.2.0 and a banner announcing Kimi Code Desktop. The post drew about 90,000 views in its first hour. It is the first time the company has shown the name, and it still gives no date, price, size or weights.
Update, October 22, 2026: A new anonymous model, Kibitz, appeared in Arena’s battles on October 21. AiBattle reported at 21:26 UTC that “The Kibitz model claims to be a Kimi model,” and warned that “Asking models about their identity in the Code Arena is unreliable.” Three hours later AI Tracker Bot reported that Kibitz “garbled all 11 Moonshot-only test tokens in both tries, like every Kimi model we tested.” Moonshot, Kimi and Arena have said nothing, and no Moonshot surface lists a K3.1. At 02:21 on the morning of September 19, 2026, Beijing time, the verified Kimi account on Zhihu published a post with no words in it. One hundred and thirty-five digits: 415926535897932384626433832795… and on for another hundred. Eight hours later the first English-language relay on X had decoded it. Pi is 3.14159265358979…; remove the leading 3.1 and the remainder is, digit for digit, what Kimi posted. By late morning UTC the reading was everywhere: Kimi K3.1 is coming.
This page is the record for Kimi K3.1: what Moonshot has actually said (nothing, beyond the digits), what the rumor chain said before this and when, and what Moonshot’s own release cadence implies. It sits beside our trackers for Grok 4.8, Fable 5.2 and GPT-6 Sol, and it is written the same way: every claim dated, every absence checked and named. Times are UTC unless marked Beijing.
On this page · 12 sectionsOpen
Moonshot released Kimi K3.1 Pro on October 23, 2026, 99 days after K3. The pi post of September 19 and the teaser earlier that day both pointed at the right name.
On Moonshot’s API it costs $2.00 per million input tokens and $10.00 output, a third less than K3’s $3.00 and $15.00. K3.1 Pro Highspeed runs 2 to 3 times faster for $4.00 and $20.00.
Of the earlier claims, the July report’s August date was wrong, korrine and Kibitz are still unconfirmed, and a forecast of a release before Tuesday, October 27 was right.
It is the first .1 name in Moonshot’s line and the first model of a new K3.1 series: a post-trained version of K3 with much more post-training compute.
In Moonshot’s own runs it scores 65.7 on Terminal-Bench 4.0, beside Claude Opus 5.5 at 66.4 and ahead of GPT-6 Astra at 58.2, and leads SWE Marathon at 51.7.
The size is K3’s: 2.8T total parameters, 104B active, a 1M-token context and multimodal input. Open weights are promised by November 4, 2026.
Still open: the weights, broader independent runs, and whether Kibitz on Arena was K3.1. Artificial Analysis ranks it #4 of 215 on its Intelligence Index and #1 open model on its Cyber Index.
§ 01Kimi K3.1 Pro, as released
Moonshot calls K3.1 Pro “the first model in our new K3.1 series”. It is a post-trained version of K3 with substantially larger post-training compute, so the base is the same 2.8T model. It arrived 99 days after K3, the longest gap since January, and about 15.5 hours after the teaser.
| Item | Kimi K3.1 Pro |
|---|---|
| Announced | October 23, 2026, 17:15 UTC, on X and in a tech blog |
| Base | A post-trained version of Kimi K3 |
| Parameters | 2.8T total, 104B active (mixture of experts) |
| Context window | 1M tokens |
| Input | Multimodal |
| Versions | K3.1 Pro, and K3.1 Pro Highspeed at 2 to 3 times the speed |
| API ids | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
| Available in | Kimi.com and apps, Kimi Work, Kimi Code, Kimi API |
| Open weights | Promised by November 4, 2026 |
| Enterprise | Off by default; administrators turn it on |
| Model | Input | Output | Cache hit |
|---|---|---|---|
| Kimi K3.1 Pro | $2.00 | $10.00 | $0.20 |
| Kimi K3.1 Pro Highspeed | $4.00 | $20.00 | $0.20 |
| Kimi K3 | $3.00 | $15.00 | $0.30 |
The new model costs a third less than the one it follows, on both input and output. Speed costs double.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 66.4 | 58.2 |
| SWE Marathon | 51.7 | 46.3 | 42.5 |
| Agents’ Last Exam | 43.2 | 42.6 | 32.8 |
| OfficeQA Pro v2 | 64.4 | 62.2 | 54.4 |
| Humanity’s Last Exam, with tools | 70.0 | 67.7 | 57.2 |
| OSWorld V2.1 | 79.6 | 81.8 | 75.8 |
| BenchCAD, with tools | 92.7 | 96.2 | 95.9 |
| Terminal-Bench Science 0.1 | 38.1 | 63.3 | 68.1 |
Every model ran at maximum thinking effort, and Moonshot notes that Claude models may include fallbacks. K3.1 Pro leads on long agentic work and trails badly on science terminal tasks, 38.1 against Opus 5.5’s 63.3.
The first independent score came from Artificial Analysis on October 23 at 22:51 UTC: “Kimi K3.1 Pro is the new #1 openly-available model for agentic cyber defense”. It scores 64% on the Artificial Analysis Cyber Index, behind only GPT-6 Sol, which OpenAI offers only through trusted access. Artificial Analysis has since placed it #4 of 215 models on its general Intelligence Index, with a score of 55, at $2.11 per index task and 73.5 output tokens per second; its model page labels K3.1 Pro proprietary for now, since the weights are not out. DesignArena projects it #1 among open-weight models on its Website leaderboard and tied for #3 overall, 15 Elo points ahead of K3, a projection rather than a settled rank. On the cyber number, Moonshot itself removed cyber-related training data from the open weights (see below), so the released checkpoint may score lower.
Two more numbers from the release. On Moonshot’s internal hallucination stress test, the claim-level error rate fell from 23.4% for K3 to 8.3%, and Moonshot says those rates “do not represent overall error rates in daily use”. For the open weights it removed cyber-related training data: on its internal exploit benchmark the unfiltered checkpoint solved 90.5% of tasks and the released model 81.0%.
| Claim | Who | Verdict |
|---|---|---|
| K3.1 ships in August with open weights | July 31 report | Wrong on the date; weights now promised by November 4 |
| A third-party tool labels its model K3.1 | Bilibili users, August 20 | Name right, nothing else checkable |
| korrine on Code Arena is K3.1 | Guess, August 22 | Never confirmed |
| The pi post means K3.1 | Market reading, September 19 | Right |
| Kibitz on Arena is a Moonshot model | Testers, October 21 and 22 | Not confirmed by Moonshot or Arena |
| Out before Tuesday, open source, low cost | @ChrisGPT, October 22 | Right on timing and price; weights still to come |
| The teaser shows Kimi K3.1 in Kimi Code | Kimi on X, October 23 | Right; it shipped about 15.5 hours later |
§ 02What we knew before launch
| Claim | Status |
|---|---|
| Kimi K3.1 exists | Shown on screen in Moonshot’s own teaser video (October 23, 01:41 UTC); not yet announced in words |
| The post decodes to 3.1 | True: the 135 digits equal pi’s expansion immediately after 3.1; we checked against 200 digits |
| The post is authentic | Two independent screenshots show the same digits on the account with Zhihu’s verified-organization badge; the post itself is behind Zhihu’s login wall, so we could not open it |
| Release date | Nothing from anyone; a July 31 report said August, which passed |
| Open weights | Claimed by the July 31 report; nothing from Moonshot |
| Size, price, context, benchmarks, model id | Nothing exists |
| Moonshot surfaces (model list, pricing, news index, Hugging Face, OpenRouter) | No K3.1, no kimi-k3.1 (read 13:20 UTC, September 19) |
| Kimi on X | A captionless video on October 23 shows Kimi Code with its model set to Kimi K3.1; no date, price or spec |
| Anonymous checkpoints | korrine on Code Arena (August 22) guessed as K3.1, never confirmed; adamant-ananke floated the same week |
| Kibitz on Arena (October 21) | Says it is Kimi and failed 11 of 11 Moonshot-only tokenizer tests (testers on X); Arena and Moonshot have not named it |
§ 03What Kimi posted, and what it decodes to
The post itself is the shortest primary source we have ever tracked. No caption, no image, no hashtag. The account is the one Zhihu marks 已认证机构号, a verified organization account, and the timestamp in both screenshots reads 2026-09-19 02:21, which is 18:21 UTC on September 18. We have not been able to open the post: Zhihu shows anonymous readers a verification wall and its API returns nothing for the account’s pins from outside China. What we can verify is the arithmetic. The string matches pi’s decimal expansion from the third digit onward for all 135 places, and the digit that would come next is 3.
The first English relay we can date came from @MaxForAI at 10:19 UTC: “Kimi’s official channels haven’t formally announced K3.1 yet”, with the note that the flagship on the website and docs is still K3. Seventeen minutes later @NFT_Chen posted the same reading; at 11:02 @notjazii wrote that “we might see it next week” and that “k3 was around opus 5 level, so k3.1 might even mog fable 5”, which is a guess stacked on a guess. The skeptics were in the same threads. One reply: “this is not a hint, this is a company that took a math class and wants credit”. Another: “doesn’t it mean: there is no 3.1?” Both readings fit a post with no words in it. What does not fit is the timing: a verified company account posting pi minus 3.1 at two in the morning, seven weeks after a report said K3.1 was due, and the company saying nothing afterward.
§ 04The rumor chain, dated
| Date | What was said | Source | Weight |
|---|---|---|---|
| Jul 31 | K3.1 will launch in August and stay open-weight | An AI industry note, relayed by 什么值得买 on Aug 21 | Report; the date passed |
| Aug 20 | The third-party tool workbuddy quietly relabelled its Kimi model K3.1 | Bilibili users, relayed by 什么值得买 | Third-party label, unconfirmed |
| Aug 22 | An anonymous Code Arena model, korrine, is guessed to be K3.1 because K3 tested as kivine | 机器之心, relayed by 什么值得买 on Aug 26 | Codename guess; Qwen and MiMo V3 also proposed |
| Aug 26 | A rival codename, adamant-ananke, is floated for Moonshot’s next model | Relayed by 什么值得买 | Unsourced |
| Sep 19, 02:21 Beijing | Kimi’s verified Zhihu account posts 135 digits of pi after 3.1 | Kimi on Zhihu (screenshots) | The only first-party signal |
| Sep 19, 10:19 UTC | First English decode: K3.1 is coming | @MaxForAI on X | Relay |
| Oct 21, 21:26 UTC | Kibitz, a new anonymous Arena model, says it is Kimi; the tester warns identity answers are unreliable | @AiBattle_ on X | Self-ID; GLM models also claim to be Kimi |
| Oct 22, 00:44 UTC | Kibitz garbles all 11 Moonshot-only test tokens, like every Kimi model tested | @aitrackerbot on X | Tokenizer evidence; points to Moonshot, not to a version |
| Oct 22, 08:07 UTC | K3.1 will release before Tuesday next week, open source, low cost | @ChrisGPT on X | No source named |
| Oct 23, 01:41 UTC | A video with no caption: Kimi Code on a CRT monitor, model set to Kimi K3.1, thinking on high, a Three.js prompt | Kimi on X | First-party; names the model on screen |
Read the chain for what it is. Everything before September 19 is other people naming the model: a report with a date that failed, a vendor’s label nobody confirmed, a codename argument that had three candidates. The pi post is different in kind, because it comes from the company, and different in content, because it says nothing. A teaser without a word in it commits Moonshot to nothing, and that may be the point.
Kibitz is the strongest outside signal yet, and it still is not Moonshot speaking. A model saying it is Kimi proves little: the tester who first spotted it wrote that “GLM and other models also sometimes identify as Kimi models.” The tokenizer test is harder to fake, since it reads which vocabulary the model was trained on, and it says Kibitz is a Moonshot model. It does not say which version, or that it ships soon. A second anonymous model, Kalmia, appeared the same morning on Qwen’s tokenizer.
§ 05Moonshot’s release cadence
The one thing Moonshot cannot hide is its own history. Seven numbered releases in twenty months, each on a public page.
| Model | Announced | Days since prior | Primary |
|---|---|---|---|
| Kimi K1.5 | Jan 20, 2025 | Kimi on X, arXiv 2501.12599 | |
| Kimi K2 | Jul 11, 2025 | 172 | Kimi on X, moonshotai.github.io |
| Kimi K2 Thinking | Nov 6, 2025 | 118 | Kimi on X |
| Kimi K2.5 | Jan 27, 2026 | 82 | kimi.com model page |
| Kimi K2.6 | Apr 20, 2026 | 83 | kimi.com, Hugging Face |
| Kimi K2.7 Code | Jun 12, 2026 | 53 | Kimi on X, Hugging Face |
| Kimi K3 | Jul 16, 2026 | 34 | kimi.com/news, Kimi on X |
| Kimi K3.1 | Not announced | 98 and counting (Oct 22) | None |
Two things stand out. The gaps shortened every cycle from January to July, and the current quiet stretch, 98 days as of October 22, is longer than every gap since January. That is consistent with a point release being close, and equally consistent with a company that shipped its largest model ever in July and is taking longer over the next one. Moonshot has also never used a .1 name: its habits are half-step increments (K2.5, K2.6, K2.7) and dated builds (kimi-k2-0905-preview is still in the model list). If the pi reading is right, K3.1 is a new naming convention as well as a new model.
§ 06What a K3.1 would follow
| Item | K3 |
|---|---|
| Total parameters | 2.8T (mixture of experts; 896 experts, 16 active per token) |
| Active parameters | 104B |
| Context window | 1,048,576 tokens |
| Headline benchmark | GPQA Diamond 93.5, per Moonshot’s model card |
| API price | $3.00 per million input tokens ($0.30 on a cache hit), $15.00 output, per Moonshot’s pricing page; third-party hosts on OpenRouter start at $1.70 and $8.50 |
| Weights | Open since July 27, 2026 (moonshotai/Kimi-K3) |
| Announced | July 16, 2026 |
Whatever K3.1 is, it starts from a 2.8T open-weight model with a million-token context and a $15 output price. The August rumor’s one concrete claim, that K3.1 stays open-weight, would put it in the same market as DeepSeek V4.1 Flash and GLM-5.3-Flash, both of which our DeepSeek tracker covers, and both of which shipped as documented releases, not riddles.
§ 07What would have confirmed it
- A line on kimi.com or platform.moonshot.ai. A news post, a model page or a kimi-k3.1 identifier in the model list. K3’s launch post and API arrived the same day.
- Weights. A moonshotai/Kimi-K3.1 repository on Hugging Face; K3’s weights followed its announcement by eleven days, so a weights-first release would be a first.
- A named checkpoint. Code Arena or OpenRouter attaching Moonshot’s name to Kibitz, korrine, adamant-ananke or a new codename.
- A second post with words in it. Kimi’s Zhihu, WeChat or X accounts saying the name.
- A price. Anything relative to K3’s $3.00 and $15.00.
§ 08Where CellCog stands
Our conflict, declared: we build CellCog, a platform where a business hires AI employees that keep memory and work as a team, and we choose the models underneath them. Today every tier runs Claude Opus 5.5. We do not route to Kimi. We track Moonshot’s models the way an operator does, as candidate engines, and an open-weight K3.1 strong at long-horizon agentic work is one we would test on our own infrastructure once the weights land. We had no part in this release.
§ 09What we are watching for
- The weights on Hugging Face, promised by November 4, 2026.
- More independent runs. Artificial Analysis has published its Intelligence Index (#4) and Cyber Index scores, DesignArena a projected rank, and Arena’s Agent board a preliminary #11 (read October 26); a settled Arena rank and a run on the open weights themselves are next.
- Third-party hosts on OpenRouter. It listed kimi-k3.1-pro at 17:56 UTC on October 23 at Moonshot’s $2.00 and $10.00, served only by Moonshot so far; for K3, other hosts came in below Moonshot’s own price.
- Arena naming Kibitz, or any other anonymous model as Moonshot’s.
§ 10The tracker
This page updates when facts change, not when posts get louder. As of October 24, 2026, 12:30 UTC: Kimi K3.1 Pro is released, priced and benchmarked by Moonshot, and listed on OpenRouter; its first independent scores are in (#4 of 215 on the Artificial Analysis Intelligence Index, 64% on its Cyber Index, and a preliminary #11 on Arena’s Agent board), while its weights and broader independent runs are still to come.
§ 11Update log
As of October 26, 2026, 02:56 UTC: Arena’s Agent board read in our browser lists Kimi K3.1 Pro (Max) at #11, marked Preliminary, with 7.61% net improvement (±3.00%) over 4,071 votes, a rank range of 7 to 15. Kimi still does not name Kibitz.
This is a living page; when the story moves, the update lands here.
As of September 19, 2026, 13:30 UTC: page opened. Kimi’s Zhihu post read from two independent screenshots (02:21 Beijing); the digits verified against pi; Moonshot’s model list and pricing page read in a browser at 13:20 UTC, the news index, the K3 model card and OpenRouter’s catalog at 13:10 UTC, zero mentions of K3.1; Kimi’s X timeline read at 13:05 UTC, last post September 9.
As of October 22, 2026, 09:30 UTC: Kibitz added. AiBattle’s and AI Tracker Bot’s posts read verbatim on X (21:26 UTC October 21, 00:44 UTC October 22); Kimi’s X timeline read 09:20 UTC (latest October 15, nothing on K3.1); OpenRouter’s catalog and Moonshot’s Hugging Face models read 09:20 UTC, no K3.1; Arena’s X account has posted nothing on Kibitz. Routing line corrected to Opus 5.5.
As of October 23, 2026, 02:40 UTC: Moonshot’s teaser added. Kimi’s X post read in a browser at 02:29 UTC (01:41:49 UTC, clocked from the post id), about 90,000 views; poster frame saved and read: model Kimi K3.1, Kimi Code 2.2.0, a Kimi Code Desktop banner. OpenRouter and Moonshot’s Hugging Face still show no K3.1.
As of October 23, 2026, 18:00 UTC: release flip. Kimi’s launch post on X read in a browser (17:15:24 UTC), the tech blog read in full, the pricing page read for kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and the benchmark overview image read. OpenRouter and Hugging Face carry no K3.1 yet. Rumor scorecard added.
As of October 24, 2026, 02:10 UTC: OpenRouter listing added. Its model catalog read at 02:08 UTC: moonshotai/kimi-k3.1-pro created 17:56 UTC on October 23, a 1,048,576-token context, $2.00 per million input tokens ($0.20 on a cache hit) and $10.00 output, from two Moonshot endpoints (the second at $4.00 and $20.00, the Highspeed price). Moonshot’s Hugging Face still shows no K3.1.
As of October 24, 2026, 07:00 UTC: first independent run added. Artificial Analysis’s post read via the X API at 06:53 UTC (22:51:18 UTC October 23): Kimi K3.1 Pro at 64% on its Cyber Index, the top openly available model, behind only trusted-access GPT-6 Sol. A creator’s claim that Moonshot paid for undisclosed promotion is not on this page: one account, one screenshot, and no response from Moonshot yet.
As of October 24, 2026, 12:30 UTC: Artificial Analysis’s model page read at 12:17 UTC: #4 of 215 on its Intelligence Index (55), $2.11 per index task, 73.5 output tokens per second, labelled proprietary until the weights ship. DesignArena’s post read via the X API (06:30:56 UTC): projected #1 open-weight model on its Website leaderboard, tied #3 overall. Moonshot’s X timeline shows nothing since its launch posts, including on the promotion claim.
§ 12Sources
- Kimi on Zhihu, verified organization account, post of 2026-09-19 02:21 Beijing time, read from screenshots posted by @MaxForAI (10:19 UTC) and @notjazii (11:02 UTC); @NFT_Chen (10:36 UTC); @kimmonismus (11:13 UTC). Post times computed from post ids.
- Moonshot AI, Kimi K3 announcement, July 16, 2026; Kimi K3 open weights, July 27, 2026; model list and pricing, read September 19, 2026, 13:20 UTC; moonshotai/Kimi-K3 model card.
- Release dates: Kimi K1.5, Kimi K2, K2 Thinking, K2.5, K2.6, K2.7 Code, K3.
- The rumor chain: 什么值得买, Kimi K3发布满月, August 21, 2026 (relaying the July 31 note and the August 20 workbuddy reports); Kimi K3.1疑似灰测, August 26, 2026 (relaying 机器之心’s August 22 korrine report).
- OpenRouter, kimi-k3 endpoints, read September 19, 2026 (Moonshot’s own endpoint at $3.00 and $15.00; third-party hosts from $1.70 and $8.50; no K3.1).
- Moonshot AI, Introducing Kimi K3.1 Pro, October 23, 2026; launch post on X, 17:15 UTC; Kimi API pricing, read October 23, 2026, 18:00 UTC.
- OpenRouter, kimi-k3.1-pro endpoints, read October 24, 2026, 02:08 UTC (listed 17:56 UTC October 23; Moonshot only, $2.00 and $10.00).
- Artificial Analysis, Cyber Index result for Kimi K3.1 Pro, X post, October 23, 2026, 22:51 UTC.
- Artificial Analysis, Kimi K3.1 Pro model page, read October 24, 2026, 12:17 UTC.
- DesignArena, Kimi K3.1 Pro projected rank, X post, October 24, 2026, 06:30 UTC.
Q1What is Kimi K3.1 Pro?
Moonshot’s newest model, released October 23, 2026: a post-trained version of Kimi K3 with 2.8T total and 104B active parameters, a 1M-token context and multimodal input.
Q2What is the API model id?
kimi-k3.1-pro, and kimi-k3.1-pro-highspeed for the faster version, on the Kimi API platform. Moonshot offers zero data retention to API customers.
Q3Is it better than Claude Opus 5.5?
Close, by Moonshot’s numbers. It leads Opus 5.5 on SWE Marathon and Agents’ Last Exam and trails it on Terminal-Bench 4.0, OSWorld V2.1, BenchCAD and Terminal-Bench Science. Moonshot ran the comparisons itself.
Q4When do the open weights come out?
Moonshot says the full weights will be released by November 4, 2026. Until then K3.1 Pro runs only on Moonshot’s own products and API.
Q5Was Kibitz on Arena Kimi K3.1?
Unconfirmed. Kibitz said it was Kimi and failed Moonshot-only tokenizer tests in October, but neither Moonshot nor Arena has named it.
Q6Does CellCog use Kimi?
No. Every CellCog tier runs Claude Opus 5.5. We track Moonshot’s models as candidate engines, and open weights would let us test K3.1 on our own infrastructure.