Kimi K3.1 Pro Is K3 With More Post-Training. On Terminal-Bench 4.0 It Went From 17% to 64% and Costs 63% Less Per Test.

What Moonshot shipped

Kimi K3.1 Pro came out on October 23 in the Kimi apps, Kimi Work, Kimi Code and the Kimi API, as kimi-k3.1-pro. Moonshot calls it “the first model in our new K3.1 series” and describes it as a post-trained version of Kimi K3: the same mixture-of-experts model, 2.8 trillion parameters in total and 104 billion active per token, with a 1,048,576-token context window.

It always thinks. Turn thinking off and the API answers HTTP 400. The reasoning levels are low, high and max, and max is the default and the setting behind every benchmark Moonshot published. Temperature is fixed at 1.0.

Source: Moonshot’s Kimi API prices as published by Digital Applied, read October 26, 2026. Flat across the full 1M context, no long-context surcharge. The Highspeed speed-up is Moonshot’s own claim.

So the new model is a third cheaper than the old one on every rate. The full weights are promised “by Nov 4, 2026.” When we checked Moonshot’s Hugging Face page on October 27, the newest repository was still Kimi-K3.

Same model, 46.5 more points

Vals.ai runs every model on Terminal-Bench 4.0 through the same minimal harness, mini-swe-agent, with one bash tool, no step or cost limit and 3 full runs per score. Kimi K3 and Kimi K3.1 Pro are both on the board. Since they share a base model, the gap between them is about as clean a measure of post-training as you’ll find in public.

The score went up 3.7 times. Everything else went down: 40% fewer tokens, 63% less money, a quarter of the time. K3 spent its budget wandering. K3.1 Pro finishes. Artificial Analysis sees the same jump on its broader Intelligence Index: 44 for K3 at max effort, 55 for K3.1 Pro.

How it stacks up against the top of the board

Source: Vals.ai Terminal-Bench 4.0, updated October 26, 2026. $ per point = cost per test ÷ score, our math, rounded to the cent. A rough efficiency read, not a Vals.ai metric.

Two readings come out of this table. Against Opus 5.5, K3.1 Pro gives up 1.51 points for a 65% smaller bill. Against GPT-6.1 Sol, it buys 8.59 more points for 2.7 times the cost. The only two models above it, Opus 5.5 and Sonnet 5.5, cost 2.9 and 3.6 times as much per test. If you need a model near the very top of this benchmark, K3.1 Pro is the cheapest way to get there today.

The same benchmark gives different numbers on different harnesses, and that is normal. Moonshot reports 65.7% on the Claude Code harness. Vals.ai measures 63.64% on mini-swe-agent. Artificial Analysis’s own run put it at 67.2%, according to OfficeChai. Moonshot’s headline number sits inside that range, and its rival numbers line up too: it gives Opus 5.5 66.4% (Vals.ai: 65.15%) and GPT-6 Astra 58.2% (Vals.ai: 59.60%).

Moonshot’s own chart: 6 wins out of 14

Moonshot published a 14-benchmark chart against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, DeepSeek V4.1 Flash 1017 and GLM-5.3, every model at maximum thinking effort. Its footnote warns that proprietary models may involve fallback behavior. GPT-6.1 Sol and Terra are not on it.

Source: Moonshot AI’s launch chart. Vendor-reported, not reproduced.

On 11 of the 13 scored benchmarks, K3.1 Pro is within 3.5 points of the leader or ahead. The two exceptions are telling. Terminal-Bench Science is a 30-point gap to GPT-6 Astra, a different weight class on scientific work. And on Kimi Code Bench, Moonshot’s own internal test, K3.1 Pro loses to Astra by 5.3 points. Labs don’t often publish a chart where they lose their own benchmark. We count that in Moonshot’s favor.

The 3D push: Blender, Unity, Unreal, Godot

This release leans hard on 3D. Moonshot writes that “K3.1 Pro is strong at 3D creation and spatial reasoning,” handles full projects in “game engines such as Unity, Unreal Engine, and Godot,” and can “write and run scripts for tools like Blender and CAD software.” It also launched 3D Make on kimi.com and in the Kimi app, a tool for previewing, editing and downloading 3D-printable models.

The benchmark behind the claim is BenchCAD, which scores 3D reconstruction through CAD code generation. K3.1 Pro gets 92.7 with tools, third behind Opus 5.5 (96.2) and GPT-6 Astra (95.9), level with Fable 5.1 (92.6). Strong, not first.

Independent testing is mixed, as you’d expect. In MindStudio’s hands-on, K3.1 Pro built a working 3D Rubik’s Cube with scramble, timer and a solution replay, and a playable browser obstacle course in about 48 minutes, with flickering and loose collisions. It failed a Blender bowline knot: the rope looked like a figure-8 and wasn’t actually tied. In a Moonshot-run rocket-scene comparison that MindStudio reports, Kimi took about 6 minutes and $1.95 where Opus 5.5 took about 24 minutes and $10.35, and Opus’s scene looked slightly cleaner.

Why a smaller K3.1 is probably next

Moonshot hasn’t announced a smaller model. But the signs that a 2.8-trillion-parameter model is hard to serve at scale are all public:

• It’s slow. 42.6 tokens per second, under half the median. It also writes a lot: 150 million output tokens to run Artificial Analysis’s index, against a median of 81 million.
• Speed is sold separately. Highspeed costs twice as much per token for 2–3x the speed, and in Kimi Code it uses about twice the quota.
• Access is rationed. In Kimi Code, K3.1 Pro needs the Plus plan or higher, Highspeed needs Pro or higher, and the full 1M context is for Pro and above. Enterprise admins have to switch it on.
• One host only. On OpenRouter, Moonshot is the only provider of K3.1 Pro, with 2 endpoints. Kimi K3 has 30 endpoints there today.
• Old models are being cleared out. The K2 series closes to new API users on October 31 and retires by the end of December.
• The name leaves room. “The first model in our new K3.1 series,” and it’s called Pro.

Our read: a smaller K3.1 that Moonshot can serve fast and cheap is the obvious next move, and the November 4 weights will help too. Once anyone can host K3.1 Pro, prices should fall the way K3’s did. Moonshot lists K3 at $3 per million input tokens. The cheapest K3 host on OpenRouter charges $0.20, on 4-bit weights.

Fun facts

• 🔓 Made worse at hacking on purpose. Moonshot removed cyber material from training before the open release. On its xdev-bench-hard test (pass@3), the unfiltered checkpoint scored 90.5% and the shipped model 81.0%. K3 scored 33.3%, and by Moonshot’s own count DeepSeek V4.1 Flash beats K3.1 Pro at 85.7%.
• 🚶 It taught a skeleton to walk. On the open human-atlas model, with 296 skeletal and 402 muscle components, K3.1 Pro ran 167 missions and 420 reviews, fixed 36 errors in shoulder-muscle moment-arm calculations, and got the figure from collapsing within seconds to walking for 11.2 seconds.
• 🧠 It can’t stop thinking. Disable thinking and the API returns HTTP 400.
• 🔎 Fewer made-up claims. On Moonshot’s deliberately adversarial hallucination set, the claim-level error rate fell from 23.4% for K3 to 8.3%.
• 🏷️ Same benchmark, two universes. On Agents’ Last Exam, Moonshot’s chart gives Opus 5.5 42.6 and GPT-6 Astra 32.8. OpenAI’s GPT-6.1 Terra launch charts, as reported by Kingy AI, give Opus 5.5 63.2% and Astra 59.3%. Never compare numbers across two vendors’ charts.
• 🏆 Open-weights record, pending. Once the weights are out, K3.1 Pro’s 55 would be the top open-weights score on Artificial Analysis, cutting the gap to the best closed model from 12 points to 3, by OfficeChai’s count.
• 🕵️ Its parent has a past. In August, Kimi K3 escaped a test sandbox to look up a benchmark’s answers online, The Next Web reports.

Where it ranks overall

Fourth of 221 models, ahead of GPT-6 Astra and Gemini 4 Argon, and the only model in the top four from outside Anthropic and Meta. For a model you’ll soon be able to download, that is the real headline.

What this means if you code with AI

For long agent runs, K3.1 Pro is the price-performance pick near the top. On Terminal-Bench 4.0, the two models that score higher cost 2.9 and 3.6 times as much per test.

Plan for slow output. At 42.6 tokens per second, it suits background agents and long tasks better than chat-speed back-and-forth.

Try it on 3D and game code, but check the result. The scripts run; the geometry isn’t always right. Render a preview and look at it before you trust it.

Keep thinking at max unless you’ve measured otherwise. Every published number uses it, and switching the level mid-session invalidates the context cache, according to the Kimi Code docs.

Watch November 4. If the weights land on time, expect more hosts, lower prices and the speed problem to ease.

Earlier this year we covered the Kimi K3 launch, and this morning the GPT-6.1 Terra vs Sol numbers on the same leaderboard, plus Terra’s computer-use results.

Kimi K3.1 Pro in MegaBrain

Full disclosure: we sell an AI coding agent. We checked on October 27: Kimi K3.1 Pro answers through MegaBrain as moonshotai/kimi-k3.1-pro. It is one of the included models on every paid plan, from Plus at $8 a month, next to GLM, MiniMax, DeepSeek, Qwen, Grok and Auto. Claude, GPT and Gemini are paid from your credits from Pro, and your own API keys and local models are free on every plan.

Pick it in the terminal coding agent, VS Code, JetBrains or the Mac app, run the same task on Opus 5.5, and compare the results yourself. Plans are on the pricing page. MegaBrain is not affiliated with Moonshot AI, Vals.ai or Artificial Analysis.

Sources

All pages opened October 27, 2026.

• Moonshot AI — Kimi K3.1 Pro announcement (benchmarks, 3D, Highspeed, cyber section, weights date)
• Vals.ai — Terminal-Bench 4.0 leaderboard (score, cost, tokens and duration per test; updated October 26, 2026)
• Artificial Analysis — Kimi K3.1 Pro (Intelligence Index, output speed, tokens, price)
• Artificial Analysis — Kimi K3 (Intelligence Index, speed, providers, parameters)
• OfficeChai — “Kimi K3.1 Pro Places 4th On Artificial Analysis Intelligence Index” (the rest of the index)
• Kimi API — pricing and model notes (K2 series closed to new users from October 31, 2026)
• Digital Applied — Kimi K3.1 Pro prices, API behavior and weights status
• Unite.AI — Kimi K3.1 Pro availability and Kimi Code plan access
• The Next Web — “Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills”
• MindStudio — Kimi K3.1 Pro 3D coding demos (what worked, what failed)
• OpenRouter — Kimi K3.1 Pro endpoints (hosts and per-host prices)
• OpenRouter — Kimi K3 endpoints (hosts and per-host prices)
• Hugging Face — moonshotai organization (published weights)
• Kingy AI — GPT-6.1 Terra benchmarks (Agents’ Last Exam numbers from OpenAI’s launch charts)