Kimi K3.1 Pro review, pricing and verdict

Artificial Analysis scores Kimi K3.1 Pro 46 on its Intelligence Index v5.0 and measured 32.1 output tokens per second, so it trades speed for reasoning depth. It suits teams running long coding or finance agents that can wait, and who would otherwise price Claude Opus 5.5 or GPT-6 Astra.

Kimi K3.1 Pro is Moonshot AI's reasoning model, released October 23, 2026 as a post-trained upgrade of Kimi K3. Moonshot reports 65.7% on Terminal-Bench 4.0, and the model accepts text and image input across a 1,048,576-token context window with thinking always on.

Provider: Moonshot AI · Family: Kimi

More about Moonshot AI on HokAI

Context window: 1,048,576 tokens · Max output: 1,048,576

Input modalities: text, image · Output: text, tool-calls

About Kimi K3.1 Pro

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, released on October 23, 2026 as a post-trained version of Kimi K3 with substantially more post-training compute. Moonshot calls it a 3T-class model, and Artificial Analysis lists 2.8 trillion parameters; active parameters are not stated. Moonshot AI sells it through the Kimi API as kimi-k3.1-pro, next to a Highspeed variant that Moonshot says runs 2 to 3 times faster for double the price. Inputs are text and images, outputs are text, and thinking is switched on permanently, with reasoning effort set to low, high or max (default max).

The headline numbers are Moonshot's own. On Terminal-Bench 4.0, which tests coding agents on resource-constrained terminal tasks, Moonshot reports 65.7% with the Claude Code scaffold, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On its internal Kimi Code Bench 3.0 Long-horizon it scores 61.2% against K3's 55.95%. On BenchCAD, a CAD-code reconstruction test, it reaches 92.7% with tools, behind Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%, and 64.4 on OfficeQA Pro v2, which Moonshot describes as the leading result there. The full results table in the launch post is an image, so only figures the post states in text are used here. Independent data is thin: Artificial Analysis scores the model 46 on its Intelligence Index v5.0, spent 110M output tokens running it, and measured 32.1 output tokens per second with 3.10 seconds to the first token. That index was rebased, so do not set the 46 beside scores recorded under an earlier version.

The API takes a 1,048,576-token context window, and max_completion_tokens defaults to 131,072 and can be raised to 1,048,576. Billing is flat per token with no long-context tier, and context caching is automatic with a 5-minute or 1-hour TTL (time to live). It supports strict JSON-schema output, partial mode, tool_choice, dynamic tool loading and vision input through base64 or uploaded file ids; public image URLs are not accepted. Sampling is fixed, so temperature and top_p cannot be tuned. Moonshot says Zero Data Retention is available for API customers. Its pricing page lists the exact rates, and the first top-up on the API platform switches the model on for an account.

Moonshot's finance and safety claims are specific but self-reported. Against K3 it reports gains of 7.2 points on internal financial modeling, 11.6 on financial research, and about 4.2 times the business-operation score. On a factual-accuracy stress test, the claim-level hallucination rate fell from about 23% for K3 to under 9%; Moonshot's two pages give 6.85% and 8.3% for the new figure, and it says the test is built to provoke errors, so these are not everyday rates. On cybersecurity, Moonshot removed cyber-related data from mid-training and reinforcement-learning stages, and its xdev-bench-hard pass@3 dropped from 90.5% for a checkpoint without those controls to 81.0%. It says the result still trails frontier proprietary models that have been public for at least three months.

The model fits long-horizon coding agents, finance and spreadsheet workflows over large document sets, and 3D or CAD scripting in tools such as Blender. It is a poor fit for anything latency-sensitive, for audio, and for self-hosting until Moonshot actually publishes the weights it promised for November 4, 2026. Teams that want downloadable weights today should look at Kimi K3, and teams that need the highest published CAD or terminal scores should compare Claude Opus 5.5 and Claude Fable 5.1 directly. DeepSeek V4.1 Flash 1017 and GLM-5.3 are the other two models Moonshot benchmarked against in its launch post.

It is available on Kimi.com and the Kimi mobile apps (iOS, Android, HarmonyOS), in the Kimi Work desktop app version 3.1.0 or later, in Kimi Code through the /model command, and through the API. Kimi Enterprise administrators must switch it on, because access is off by default. The consumer chat app is listed on HokAI as Kimi. Moonshot also says K2 series models close to new users on October 31, 2026 and retire by the end of December 2026, and points those users to K3.1 Pro; the earlier Kimi K2.7 Code HighSpeed falls in that group. For a rival at the same listed input and output token prices, see GPT-6.1 Sol, and the rest of the lineup sits on the Moonshot AI provider page. HokAI's guides Kimi vs ChatGPT, Which LLM Should You Use?, Which Claude Model Should You Use? and The Cheapest LLM APIs in 2026 predate this release, so read them as the baseline this model is measured against.

Pricing

Read from platform.kimi.ai/docs/pricing/chat on 2026-10-30. kimi-k3.1-pro bills $2.00 per 1M input tokens, $10.00 per 1M output tokens and $0.20 per 1M cache-hit input, flat across the 1,048,576-token window with no long-context tier. Cache writes cost $2.00 (5-minute TTL) or $4.00 (1-hour TTL) per 1M tokens. The Highspeed variant bills $4.00 input and $20.00 output. Prices exclude taxes. Moonshot's guide says access starts after a successful top-up of at least $1. Weights are promised by November 4, 2026, after which local inference would cost hardware only.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Always-on reasoning with three effort levels: Thinking cannot be turned off, and reasoning_effort accepts low, high or max, with max as the default.
• 1M-token context: A 1M-token window, with max_completion_tokens defaulting to 131,072 and adjustable up to the full window.
• Automatic context caching: Repeated prefixes are cached without setup, with a 5-minute default TTL or an optional 1-hour TTL, and cache hits bill at one tenth of fresh input.
• Structured output and tool control: Strict JSON-schema output, partial mode, tool_choice and dynamic tool loading through a system message.
• Image input and 3D scripting: Reads images sent as base64 or file ids and writes Blender and CAD scripts directly, with 92.7% on BenchCAD in Moonshot's tool-assisted run.

Pros

• Costs half the HokAI-listed token price of Anthropic's Opus model, per input and output token, while trailing it by 0.7 points on Moonshot's own Terminal-Bench 4.0 table.
• Moonshot has committed to release the full weights by November 4, 2026, which would allow self-hosting and fine-tuning if it delivers.
• Moonshot published its cyber-capability evaluation and the training-data removal it applied, with before and after scores.
• Moonshot reports clear finance gains over K3 on internal tests, alongside an OfficeQA Pro v2 score that Moonshot reports as the leading result.

Cons

• Slow in Artificial Analysis's measurement, and the Highspeed variant that Moonshot says is 2 to 3 times faster costs double.
• Headline scores are vendor-run, so a gap of a point or two against Opus or GPT-6 Astra is not settled.
• Hosted-only until the promised weights appear, so the open-weight appeal is a commitment, not a download.
• No audio input or output, and Moonshot's built-in web search is on hold.
• No SOC 2 or HIPAA statement was found on the pages read, which matters for regulated teams.

Benchmarks

• AA Intelligence Index: 46 cited: Artificial Analysis · 30 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $1.54/M cited: Artificial Analysis · 30 Oct 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 32 tok/s cited: Artificial Analysis · 30 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Kimi K3.1 Pro Official Page