Kimi K3.1 Pro Is Fourth on the Index at a Third Off K3's Price. The Open-Weight Title Waits on Weights Due November 4.
Moonshot shipped Kimi K3.1 Pro on October 23 at $2 in and $10 out per million tokens, a third below K3, and Artificial Analysis scores it 55, fourth on its Intelligence Index. The catches: it is a closed API until the weights land by November 4, AA calls it notably slow and very verbose, Moonshot stripped cyber training data for the open-weight release, so the API build and the download may not match, and the Highspeed variant doubles the price. We read Moonshot's post, the pricing page, AA's model page and launch analysis, and Vercel's docs.
By T.W. Ghost
Moonshot shipped Kimi K3.1 Pro today, October 23, 2026, three months after Kimi K3 arrived as the largest open model ever released. K3.1 Pro is a post-trained K3 with, in Moonshot's words, larger post-training compute: a 1M-token context, text and image input, live today on Kimi.ai, Kimi Work, Kimi Code, and the API, with a Highspeed variant beside it. The price is $2 in and $10 out per million tokens, a third below K3's $3 and $15.
Artificial Analysis scored it the same day at 55 on its Intelligence Index, fourth of 215 models, and wrote the sentence that will be quoted everywhere: K3.1 Pro "will be the leading open-weights model once its weights are published." Read the tense. Today it is a closed API. Moonshot promises full weights by November 4, and the open-weight crown is a promise with a date, not a download.
We read Moonshot's launch post, including its cybersecurity section, the platform pricing page, AA's model page and launch analysis, and Vercel's Gateway and Token Pass docs. Here is what held up.
What Shipped, and What Did Not
| Kimi K3.1 Pro | |
|---|---|
| Released | October 23, 2026, on Kimi.ai, Kimi Work, Kimi Code, and the API; Highspeed variant the same day |
| Lineage | Post-trained K3 with larger post-training compute (Moonshot) |
| API price | $2 in / $10 out per 1M tokens; cache hits $0.20; Highspeed $4 / $20 |
| Context and modalities | 1M tokens; text and image in; text out |
| Weights | Promised "by Nov 4, 2026"; not published today; license not published |
| Enterprise | Kimi Enterprise access off by default |
| Also on | Vercel AI Gateway (K3.1 Pro and Highspeed) from launch day |
| AA Intelligence Index | 55, fourth of 215 |
The price cut is real. $2 and $10 against K3's $3 and $15 is a 33% reduction on both sides, with cache hits at $0.20. Highspeed, which Moonshot says runs two to three times faster than Pro, costs exactly double at $4 and $20. The speed multiple is Moonshot's; AA has not measured Highspeed.
The weights are a date. Moonshot's post says full weights by November 4. No license terms are published, and no Hugging Face repo exists yet. Every "best open model" sentence written this week, including AA's, is written in the future tense, and ours will be too.
The Benchmarks, Sorted by Who Reported Them
House rules: independent numbers first, stated as fact with the source; Moonshot's numbers labeled as Moonshot's, with the harness.
Artificial Analysis, October 23. K3.1 Pro scores 55, fourth of 215 on the Intelligence Index, with what AA describes as large agentic gains over K3. On AA's own Terminal-Bench 4.0 run it scores 67%, the highest figure on AA's board. On GDPval-AA, the real-economic-work Elo board, it sits third at 1,846. And AA attaches two adjectives it does not hand out lightly: notably slow and very verbose.
| Metric (Artificial Analysis, Oct 23) | Kimi K3.1 Pro |
|---|---|
| Intelligence Index | 55, #4 of 215 |
| Terminal-Bench 4.0 (AA's run) | 67%, highest on AA's board |
| GDPval-AA (Elo) | 1,846, third |
| Output speed | 57.5 tokens/s |
| Output tokens to run the index | 150M |
| Cost per Intelligence Index task | about $2.11 on the model page; about $1.98 in AA's launch article |
| List price | $2 / $10 per 1M |
Two numbers in that table disagree with each other, and both are AA's. The launch article prints about $1.98 per index task; the live model page prints about $2.11. We print both with their sources and treat the model page as the live figure.
Now the vendor table. These are Moonshot's self-reported numbers from the launch post.
| Benchmark (Moonshot's table) | Kimi K3.1 Pro |
|---|---|
| Terminal-Bench 4.0 (Claude Code harness) | 65.7% |
| Kimi Code Bench 3.0, long-horizon (internal) | 61.2% |
| Kimi Hallu Bench Hard, claim-level hallucination (internal stress test) | 8.3% |
| xdev-bench-hard, pass@3, before cyber-data removal | 90.5% |
| xdev-bench-hard, pass@3, final K3.1 Pro | 81.0% |
Three honest observations:
- Vendor and independent Terminal-Bench agree for once, in the wrong direction. Moonshot reports 65.7% in the Claude Code harness; AA's own run scores 67%. The vendor number is lower than the independent one, which is rare enough to note. Both are Terminal-Bench 4.0; do not mix either with the 2.1 figures other labs still print.
- Two of the three internal benchmarks are Moonshot's own. Kimi Code Bench and Kimi Hallu Bench are in-house tests with no outside runs. They tell you what Moonshot optimized for, not where the model stands.
- Slow and verbose is the per-task catch. At 57.5 tokens per second and 150M output tokens to run the index, K3.1 Pro is a model that thinks at length. AA's roughly $2 per index task is the result of a low per-token price multiplied by a large token count. If your workload is latency-sensitive, the Highspeed variant exists for a reason, and it costs double.
The Cyber Strip, and Which Model You Will Get on November 4
Moonshot's cybersecurity section is the most interesting paragraph in the launch. Without cyber-data removal and a reinforcement-learning filter, the model scored 90.5% pass@3 on xdev-bench-hard; the final K3.1 Pro scores 81.0%. Moonshot says the removal was done "for the open-weight release."
Read that carefully. The sentence describes a capability reduction made for the weights that ship on November 4. It does not say whether the API you call today is the stripped build or the unstripped one. If the API serves the 90.5% model and the download is the 81.0% model, then AA's 55 describes something you will not be able to self-host. If both are the same build, Moonshot has not said so. We print the question, not an answer.
Where You Can Run It
Moonshot's API and the Kimi apps from today. Vercel AI Gateway added K3.1 Pro and K3.1 Pro Highspeed the same day. Kimi Enterprise customers have access switched off by default. One thing it is not on: Vercel's Token Pass, whose featured models as of today are Kimi K3, GLM 5.3, and DeepSeek V4.1 Flash. There is no K3.1 Pro Token Pass rate, and anyone quoting one is inventing it.
The gateway-versus-Moonshot price for K3.1 Pro was not checked for this post. An X post circulating a "$1.95 versus $10.35 per task" comparison against Claude Opus 5.5 is unverified, so it stays out.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Oct 23, 2026; $2 / $10 per 1M; Highspeed $4 / $20; 1M context | Verified (Moonshot post and pricing) |
| 55 on the AA Intelligence Index, fourth of 215 | Verified (Artificial Analysis, Oct 23) |
| "Leading open-weights model" | Verified as AA's future-tense sentence; today the weights are not published |
| Terminal-Bench 4.0 67% (AA) and 65.7% (Moonshot, Claude Code harness) | Verified, two harnesses, both labeled |
| Cost per AA index task | Verified with a discrepancy: $2.11 on the model page, $1.98 in the article |
| Cyber data stripped "for the open-weight release" | Verified (Moonshot); whether the API build matches the download is Unverified |
| License for the Nov 4 weights | Unverified (not published) |
| Highspeed is two to three times faster | Vendor claim; not measured by AA |
| On Vercel Token Pass | Contradicted (Token Pass lists K3, GLM 5.3, and DeepSeek V4.1 Flash only) |
Who Should Use It
Use Kimi K3.1 Pro if you run agentic coding or long-horizon knowledge work and want a fourth-place index score at a mid-tier price; AA's Terminal-Bench 4.0 and GDPval-AA placements are the strongest independent case any Chinese lab has had this year. Re-run your K3 evals rather than reusing K3 prompts; this is a different post-training.
Skip it for latency-sensitive loops unless you pay for Highspeed, and skip any plan that depends on self-hosting until November 4 delivers weights, a license, and an answer to whether the download is the model AA scored.
The frontier is elsewhere: three closed models sit above K3.1 Pro on AA's index, and it sits above every open model on the board that you can actually download, including DeepSeek V4.1 Flash 1017 at 45. Whether that ranking holds on November 4 depends on what ships. Our model comparison has the current standings side by side.
Sources
- Moonshot: Kimi K3.1 Pro
- Moonshot platform docs: K3.1 Pro quickstart
- Artificial Analysis: Kimi K3.1 Pro model page
- Artificial Analysis: Kimi K3.1 Pro launch analysis
- Vercel changelog: Kimi K3.1 Pro and Highspeed on AI Gateway
- Vercel docs: Token Pass
*Not sure whether a fourth-place model at a third off is the right trade for your work? Take the free 2-minute quiz and get matched. Then read about the open model it will have to beat on November 4, and the same-name swap that got it to 45: DeepSeek V4.1 Flash 1017.*