Kimi K3.1 Pro Review: $2/$10 Open-Weight Contender

What Changed From K3

Moonshot did not train a new base model. Kimi K3.1 Pro is a post-trained version of K3 — the same 2.8-trillion-parameter mixture-of-experts system, the same 3-trillion-parameter class, the same 1,048,576-token context window, and the same native image and video input. What changed is how much compute went into post-training, and what that bought on agentic and professional work.

Three differences matter in practice:

1. Price. Every API line is exactly one-third cheaper than K3's, which we break down below.
2. A reasoning dial. K3 launched with maximum reasoning only. K3.1 Pro exposes reasoning_effort with low, high, and max settings, defaulting to max. The model always reasons; the dial changes how much.
3. A Highspeed SKU. kimi-k3.1-pro-highspeed produces output roughly two to three times faster at twice the price, while keeping the same $0.20 cached-input rate.

The model is available through Kimi's web app, mobile app, Kimi Work desktop app, Kimi Code CLI, and the Kimi API on platform.kimi.ai under the IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot also launched Hosted Agents, a managed runtime for long-running and multi-agent tasks.

Moonshot's launch materials lean on professional workflows: 3D scene construction in Unreal Engine, CAD-style modeling in Blender and FreeCAD (including a working turbofan engine model), and financial research workbooks. Those are vendor demonstrations, not independent tests.

Benchmarks: The Agentic Jump

Artificial Analysis places K3.1 Pro fourth on its Intelligence Index at 55, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder (56). It is the highest-scoring model slated for open weights — though the weights themselves have not shipped yet.

The agentic column is where the generation gap shows: GDPval-AA rises 312 Elo points over K3 in the same harness, and AutomationBench — agentic SaaS workflows — jumps 15 points.

Moonshot's own launch chart, run at maximum thinking effort against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, DeepSeek V4.1 Flash, and GLM-5.3, tells a similar story with sharper edges:

• Terminal-Bench 4.0: 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%).
• First place on SWE Marathon (51.7), OfficeQA Pro v2 (64.4), Humanity's Last Exam with tools (70.0), GDP.pdf, ZeroBench, and Agents' Last Exam.
• Terminal-Bench Science 0.1: 38.1, against GPT-6 Astra's 68.1 — a genuine gap for scientific computing in a terminal.
• Kimi Code Bench 3.0 long-horizon: 61.2, behind GPT-6 Astra's 66.5 on Moonshot's own coding benchmark (up from K3's 55.95).
• OSWorld V2.1 computer use: 79.6, slightly behind both Claude models.

Two caveats belong next to those numbers. First, the launch-chart results are vendor runs, and some benchmarks are internal — treat them as directional until independent reproductions land. Second, harnesses differ: Artificial Analysis's own Terminal-Bench run put K3.1 Pro at 67.0%, and independent commentary notes that scores produced at different effort levels are not strictly comparable.

On factual reliability, the picture improves but does not become clean. Artificial Analysis's Omniscience index shows the hallucination rate falling from 53% to 44%. Moonshot's internal Hallu Bench Hard — deliberately weighted toward questions models get wrong — puts single-statement errors at 8.3%, down from 23.4%, with the share of answers containing at least one error at roughly 42% of K3's. Moonshot itself says that stress test is not a daily-use error rate.

Pricing: 33% Cheaper per Token

The price cut is the simplest part of the release: every line is exactly one-third lower than K3's, per Kimi's API pricing page checked on October 24, 2026.

Pricing is flat across the context window — no long-context surcharge, which matters at a million tokens. Against the models it competes with, the standard model costs half of Claude Opus 5.5's $4/$20 and one-fifth of GPT-6 Astra's $10/$50.

Consumer access runs through Kimi's membership tiers: Free, Plus at $15 per month, Pro at $31 (the first tier with Highspeed), Max at $79, and Ultra at $159, using annual billing at effective monthly rates. Max and Ultra add million-token chat sessions.

The Cost-per-Task Catch

A 33% lower rate card does not produce a 33% lower bill. On Artificial Analysis's Intelligence Index tasks, K3.1 Pro emits about 55,000 output tokens per task — 14% more than K3's 48,000 — which puts cost per task at $1.98, essentially level with K3's $2.00.

That is not a reason to avoid the model; it is a reason to measure the right number. Per-token pricing favors K3.1 Pro on every workload, but the advantage shrinks on verbose, long-output tasks unless you use the effort dial. Set reasoning_effort to low or high for routine work and reserve max for hard problems — a setting K3 never offered.

Open Weights by November 4

Moonshot says the full weights will be published by November 4, 2026. When they land, K3.1 Pro becomes the highest-scoring open-weight model on the Artificial Analysis index, narrowing the gap between the best proprietary and best open system from 12 points to 3.

The release will not be a raw dump. Moonshot says it removed cyber-offense material during mid-training and reinforcement learning; on its internal xdev-bench-hard vulnerability-research test, the filtered model scores 81.0% versus 90.5% for the unfiltered version. Its own framing is that targeted data removal reduces offense capability, while general coding strength can still transfer to security work — a caveat worth reading before deploying the weights in sensitive environments.

Until the weights ship, teams that want managed long-running agents can use Moonshot's Hosted Agents service instead of standing up their own runtime.

Limitations

• Vendor-run benchmarks. The strongest claims come from Moonshot's launch chart; independent reproductions exist for the Artificial Analysis index metrics only.
• Scientific computing. Terminal-Bench Science at 38.1 is less than 60% of GPT-6 Astra's score.
• Long-horizon Kimi Code work. 61.2 on Moonshot's own benchmark still trails GPT-6 Astra.
• Weights are a promise, not a product. Nothing can be self-hosted until November 4 at the earliest.
• Verbosity. More output tokens per task erodes the per-token price advantage.

Who Should Use Kimi K3.1 Pro

• Agentic coding teams running terminal-heavy, long-horizon tasks: this is the strongest open-weight option on Terminal-Bench 4.0 short of Claude Opus 5.5.
• Teams priced out of Opus 5.5 but unwilling to drop to a smaller model: half the per-token rate for near-frontier index scores.
• Open-weight infrastructure planners who can wait for the November 4 release and want the highest-scoring weights available.
• Not yet for scientific computing pipelines, teams that need weights today, or workloads where per-task cost dominates and output volume cannot be controlled.

Final Verdict

Kimi K3.1 Pro is the most consequential open-weight release of the year, and the least dramatic one: no new architecture, no parameter-count jump, just a post-training cycle that bought real agentic gains and a one-third price cut. Fourth on the Intelligence Index at 55, first on several agentic benchmarks in vendor runs, and three points from the frontier once the weights land.

The honest summary is two numbers: 33% cheaper per token, and roughly the same cost per finished task. Which one you experience depends on how much output your workload generates and whether you use the effort dial. For new agentic builds, K3.1 Pro is now the default Kimi — K3 is the worse deal on both price and reported quality. For scientific computing in a terminal, it is not the model yet.

Related Articles

• Kimi K3 Review
• Kimi K3.1 Pro vs Claude Opus 5.5
• Kimi K3 vs DeepSeek
• DeepSeek V4.1-Flash Review
• GLM-5.3 Review
• GPT-6 Astra Review
• Claude Sonnet 5.5 Review
• Best Chinese AI Models 2026
• Best AI Coding Tools 2026