Moonshot AI released Kimi K3.1 Pro on October 23, 2026, and within a day it did something no open-weight model had done before: it landed fourth on the Artificial Analysis Intelligence Index with a score of 55, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56). It costs $2 per million input tokens and $10 per million output tokens — half the list price of Claude Opus 5.5 — and Moonshot promises to publish the full model weights by November 4, 2026.
If you’re a freelancer, a small agency, or a solo builder paying for AI out of your own pocket, this is the rare model launch that changes your actual monthly bill. But there’s a catch you need to hear before the headline: the weights aren’t out yet. Right now Kimi K3.1 Pro is an API product with an open-weight promise attached. What you do this week depends on whether you need the API today or the weights tomorrow. This guide gives you the verified numbers, the honest gaps, and a decision framework — plus real cost math for the kind of work freelancers actually run.
What Kimi K3.1 Pro actually is
Kimi K3.1 Pro is the first model in Moonshot AI’s K3.1 series, post-trained from Kimi K3 — the 2.8-trillion-parameter mixture-of-experts system Moonshot fully open-weighted in July 2026, then the largest open-weight model in existence. K3.1 Pro keeps the 3-trillion-parameter class (2.8 trillion total, 104 billion active parameters) and a 1-million-token context window, and it accepts text and image inputs.
It’s available right now in three places: the Kimi.ai chat app and mobile apps, Kimi Work 3.1.0 and Kimi Code, and the Kimi API under the model names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The Highspeed variant runs two to three times faster at twice the standard token price, except cache-hit input tokens, which cost the same in both tiers. One practical note for trialing it: reports from the launch coverage say the Kimi API quickstart requires a minimum $1 top-up, so treat it as a very cheap trial rather than a free one.
One more thing worth knowing up front: Moonshot deliberately removed cybersecurity-related data from the model’s post-training dataset. That means the open weights, when they arrive, will be intentionally capped on offensive security capabilities. If your work involves security tooling, plan around that limitation rather than discovering it mid-project.
The benchmarks: what Moonshot claims vs what outsiders measured
Every benchmark number in this section needs a label, because launch-week numbers come from two very different sources: Moonshot’s own runs, and one independent index.
Moonshot’s own numbers (vendor-reported):
| Benchmark | Kimi K3.1 Pro | Closest rival |
|---|---|---|
| Terminal-Bench 4.0 (coding agents) | 65.7% | Claude Opus 5.5: 66.4%, GPT-6 Astra: 58.2% |
| BenchCAD (3D modeling) | 92.7% | Within ~4 points of GPT-6 Astra and Opus 5.5 |
| Kimi Code Bench 3.0 (long-horizon coding) | 61.2% | K3: 55.95% |
| SWE-Marathon | 51.7% | Opus 5.5: 46.3% |
| Hallucination rate | 8.3% (down from 23.4%) | — |
| GDPval-AA (agentic Elo) | 1,846 | Opus 5.5: 1,866, Muse Wonder Full: 1,857 |
The independent number: Artificial Analysis published an Intelligence Index score of 55 for Kimi K3.1 Pro on launch day — an 11-point jump over Kimi K3, and the first outside measurement to set against Moonshot’s table. If the weights ship as promised, Artificial Analysis notes the gap between the leading open-weight and proprietary systems would shrink from 12 points to 3.
The honest gaps, which the headlines skip:
- Terminal-Bench Science is weak. K3.1 Pro scores 38.1 there, against 63.3 for Opus 5.5 and 68.1 for GPT-6 Astra. If your work is scientific computing or research-grade analysis, the headline coding numbers don’t transfer.
- Several Moonshot benchmarks aren’t public. Kimi Code Bench, Hallu Bench Hard, and xdev-bench-hard are internal tests, so outsiders can’t reproduce them yet.
- Rival scores in Moonshot’s table mix sources — leaderboard maxima, system cards, and Moonshot’s own runs — so head-to-head rows aren’t apples-to-apples.
- The license for the open weights hasn’t been stated yet. Until the November 4 release, nobody can confirm what commercial use the weights will allow.
The fair summary: this is a genuinely strong coding and agentic model with one independent score backing the hype — and with real, named weak spots. That’s a much better position than most launch-week models, and a much worse position than “proven.”
The cost math: what Kimi K3.1 Pro saves a freelancer
Here’s where the launch turns into money. The rate card is $2 per million input tokens and $10 per million output tokens, with cache-hit inputs at $0.20 per million. Claude Opus 5.5’s list price is $4/$20 — double on both counts. And Artificial Analysis measured K3.1 Pro’s average Intelligence Index task cost at $2.11 per task, while Moonshot lists roughly $1.98 per task.
But sticker price isn’t task price. K3.1 Pro uses about 55,000 output tokens per Intelligence Index task — 14% more than K3 — because its default reasoning effort is “max.” Heavier reasoning burns more tokens, which eats into the rate advantage. The question is whether it still wins per finished job. Let’s work three realistic freelancer scenarios at the standard tier (not Highspeed):
Scenario 1: The weekly code review. You feed a 40,000-token codebase context (input) and get back a 6,000-token review with fixes. Cost: 40,000 × $2/1M = $0.08 input, plus 6,000 × $10/1M = $0.06 output. Total: $0.14 per review. On Opus 5.5’s rate card, the same job is $0.28. Run 20 reviews a month and you’ve spent $2.80 instead of $5.60.
Scenario 2: The long-document job. A client hands you a 400,000-token research dump to summarize into a 15,000-token brief, with the system prompt cached across repeat runs. Input: 400,000 × $2/1M = $0.80 (or $0.08 on cache hits for repeat passes). Output: 15,000 × $10/1M = $0.15. Total: about $0.95 per brief — under a dollar for work you’d bill at $150+.
Scenario 3: The agentic coding session. An agent loop that runs 10 tool-calling rounds, burning 200,000 input tokens and 80,000 output tokens across the session. Cost: $0.40 + $0.80 = $1.20 per session. At Opus 5.5 rates that’s $2.40. Over 100 sessions a month, the difference is $120 — real money for a solo operator, and that’s before the 1-million-token context window saves you from chunking and re-feeding documents.
The Highspeed tier doubles those figures for 2–3x the speed. Use it for interactive work — a live coding assistant, a client-facing chat — where latency costs you. Use the standard tier for batch jobs, overnight runs, and anything where an extra minute doesn’t matter.
One more cost angle: the 1-million-token context window itself saves money. Every time you re-feed a document because a smaller model forgot it, you pay again. A model that holds your whole project in one request can be cheaper per finished job even at the same token price.
API now vs weights on November 4: the decision framework
This is the decision most coverage skips, and it’s the one that matters. Here’s how to choose:
Use the API this week if you:
– Run coding agents or long-document work and want to test a cheaper frontier-class model on your own tasks before committing
– Need the 1-million-token context window now, not in two weeks
– Are comfortable with a hosted API (your prompts go to Moonshot’s servers) and a Chinese provider’s terms
– Want to compare pass rates and cost per finished task against your current model with real data
Wait for the November 4 weights if you:
– Need to self-host — for client confidentiality, data-residency requirements, or offline work
– Want to fine-tune the model on your own data
– Need to verify the license terms before building a product on it
– Run security-sensitive workloads where you need to know exactly what’s in the model
Skip it for now if you:
– Do scientific or research-grade analysis (remember Terminal-Bench Science: 38.1)
– Need offensive security capabilities (deliberately removed from training)
– Can’t tolerate a provider whose weights, license, and long-term support are still unproven
A practical middle path, and the one I’d take: spend the $1 top-up this week, run your five most common jobs through the API at standard tier, and record three numbers per job — cost, time, and whether the output passed your own quality check. By November 4 you’ll know from your own data whether the weights are worth waiting for, instead of deciding from a benchmark table.
What the community actually said
Launch-day reaction was unusually substantive. The release drew roughly 850 points on Hacker News with nearly 500 comments by Friday evening, as reported by CGTN. The most-shared observation was about the model’s temperament — one highly visible commenter wrote: “Kimi is so much more blunt and argumentative than USA models. It’s so funny, it will say things like, ‘We’ve already been over this 3 times and you refuse to accept my advice, asking the same question another way won’t change the answer.'”
That bluntness is more than a personality quirk for freelancers: a model that pushes back on a bad instruction instead of agreeably executing it can save you from shipping a client’s mistake. It’s also a warning — if you pipe Kimi’s raw output to clients, expect a tone that reads very differently from the polished agreeableness of US models. Budget a human pass before anything client-facing ships.
Skepticism showed up too, and it’s healthy: commenters debated whether the Artificial Analysis score reflected “benchmaxxing” — optimizing for benchmarks — and whether Chinese labs had genuinely reached the frontier tier or just the appearance of it. The honest answer is that one independent index plus vendor tables isn’t proof, but it’s also not nothing. The November 4 weight release is the moment that debate gets settled, because the community will be able to run its own tests.
How Kimi K3.1 Pro fits the 2026 model market
For context on pricing, our earlier comparison of Gemini 4 Argon, GPT-6 Astra, and Claude’s flagship pricing lays out the frontier rate cards — Kimi slots in at half of Opus 5.5’s price with an independent score 3 points behind it. And if you’re weighing Claude specifically, see our coverage of Claude Opus 5.5’s cheaper launch for what changed on Anthropic’s side.
The bigger picture: Kimi K3.1 Pro is the third open-weight pressure point in a month, alongside other open releases, and it arrives as closed labs keep raising prices for their top tiers. The pattern for freelancers is consistent — every time a strong open model ships, the practical move is the same: test it on your own work, keep the incumbent for the jobs where it wins, and let the price war pay for your experimentation.
The bottom line
Kimi K3.1 Pro is the most capable soon-to-be-open model ever measured, at half the price of the model it trails by 3 points — with a weak science score, a capped security profile, and weights that are still a promise until November 4. For freelancers, the smart play isn’t switching everything today. It’s spending a dollar, running your real jobs through the API this week, and letting your own pass rates and per-task costs decide. If the weights land as promised, you’ll already know exactly what to do with them.
Frequently Asked Questions
Is Kimi K3.1 Pro actually open source?
Not yet. Moonshot has released Kimi K3.1 Pro as a hosted API model and promises to publish the full weights by November 4, 2026. Until the weights appear, “open-weight” is a promise, not a fact — and the license terms for commercial use haven’t been stated. Treat November 4 as the verification date.
How much does Kimi K3.1 Pro cost?
The API costs $2 per million input tokens and $10 per million output tokens, with cache-hit inputs at $0.20 per million. The Highspeed variant runs 2–3x faster at twice the standard price (cache hits cost the same). That’s half the list price of Claude Opus 5.5 ($4/$20). Artificial Analysis measured an average task cost of $2.11 on its Intelligence Index workload.
How good is Kimi K3.1 Pro compared to Claude Opus 5.5?
On the independent Artificial Analysis Intelligence Index, Kimi K3.1 Pro scores 55 versus Opus 5.5’s 58 — fourth overall. On Moonshot’s own Terminal-Bench 4.0 run, K3.1 Pro hits 65.7% versus Opus 5.5’s 66.4%. But K3.1 Pro trails badly on Terminal-Bench Science (38.1 vs 63.3), so the “near-frontier” claim holds for coding and agentic work, not for scientific analysis.
Can I self-host Kimi K3.1 Pro?
Not until the weights are published, promised by November 4, 2026. Right now the only way to run it is through the Kimi API, Kimi.ai, Kimi Work, or Kimi Code. Self-hosting a 3-trillion-parameter-class model will also need serious GPU infrastructure — plan for that before assuming the weights alone solve your hosting question.
Is Kimi K3.1 Pro good for freelancers?
For coding, document work, and agentic tasks, yes — the price-to-capability ratio is the best in its class right now, and the 1-million-token context window cuts re-feeding costs. For scientific analysis, security work, or client-confidential workloads that can’t touch a hosted API, wait for the weights and the license, or stick with your current model.
What’s the catch with Kimi K3.1 Pro?
Three catches: the weights aren’t out yet (November 4 promise), the license is unstated, and Moonshot deliberately removed cybersecurity data from post-training, capping the model’s offensive security abilities. Also note that several of Moonshot’s own benchmarks are internal tests that outsiders can’t reproduce yet, so independent verification is still thin.