Moonshot's Kimi K3.1 Pro Nears Frontier at Half Opus Pricing

Moonshot AI released Kimi K3.1 Pro on Oct. 23 and committed to publishing the full model weights by Nov. 4 — a rare near-frontier model that research groups will be able to run on their own hardware for data they would rather keep off vendor servers.

Independent scoring puts the model close to the top of a crowded field. Artificial Analysis gives K3.1 Pro a 55 on its Intelligence Index, fourth among 215 models. Anthropic's Claude Opus 5.5, released Sept. 22, leads at 58; Claude Sonnet 5.5 follows at 56; Meta's Muse Wonder Full, introduced Oct. 12, also scores 56.

Claude Fable 5.1, OpenAI's GPT-6 Astra and Google's Gemini 4 Argon each score 53. Google unveiled Argon on Sept. 30, claiming frontier-level benchmark performance, but Artificial Analysis ranks it behind both GPT-6 Astra and Opus 5.5.

What does it cost to run?

Moonshot charges $2 per million input tokens and $10 per million output tokens, down from $3 and $15 for Kimi K3, which launched in July. Those rates match GPT-6 Astra's successor-tier GPT-6.1 Sol and are half Opus 5.5's price, as software developer and AI blogger Simon Willison noted. For a lab running heavy API workloads, that difference compounds quickly.

A Highspeed variant, which Moonshot says runs two to three times faster, costs $4 per million input tokens and $20 per million output tokens. That matches Opus 5.5 and doubles GPT-6.1 Sol's per-token price. Cached input rates remain unchanged for the faster version.

Artificial Analysis ran an unusually heavy evaluation — 150 million output tokens against a median of 87 million — and measured an output speed of 77 tokens per second. Its weighted average cost per Intelligence Index task came to $2.11 for K3.1 Pro, against $5.98 for Opus 5.5, $7.63 for Fable 5.1 and $3.26 for GPT-6 Astra. Gemini 4 Argon cost $1.99 per task at a score of 53; GPT-6.1 Sol cost $0.72 at 52.

What do the benchmarks actually show?

Moonshot's own benchmark sheet compares K3.1 Pro against Opus 5.5, Fable 5.1, GPT-6 Astra, DeepSeek V4.1 Flash and Zhipu's GLM-5.3 across 14 tests. The model posts the top score on six, including:

• SWE Marathon, a long-horizon software engineering test: 51.7 versus 46.3 for Opus 5.5
• Humanity's Last Exam with tools: 70.0 versus 67.7
• OfficeQA Pro v2, a financial-document analysis test: 64.4 versus 62.2
• Terminal-Bench 4.0: 65.7, behind Opus 5.5 at 66.4 and ahead of GPT-6 Astra at 58.2

These are vendor-run comparisons, and the methodology deserves scrutiny. Rival scores on the sheet mix Moonshot's own runs, vendor system cards, blogs and public leaderboards. Moonshot notes Claude results may include fallbacks and GPT results may include cyber safeguards. Kimi Code Bench 3.0 and Kimi Hallu Bench are internal Moonshot benchmarks, and the DeepSeek score on Terminal-Bench Science comes from an older version — Moonshot's own run of the current one scored 9.3.

Where does it fall short for research workflows?

The weakest signal for research readers is Terminal-Bench Science 0.1, which tests agents on scientific research workflows in a terminal. K3.1 Pro scores 38.1, against 68.1 for GPT-6 Astra, 63.3 for Opus 5.5 and 40.0 for Fable 5.1. That is a 30-point gap to the leader in the task category closest to daily lab automation.

An August preprint introducing BenchBench-Protocol, built on 149 modifications scientists made to published wet-lab protocols, complicates the picture. Kimi K3 and GPT-5.6 Sol were statistically indistinguishable on single attempts; over 10 attempts, Kimi K3 scored higher. Claude Opus 5 led the nine models tested at 59.2%.

Independent testing also tempers the frontier claim. Peter Gostev, AI capability lead at Arena, the evaluation company formerly known as LMArena, tested K3.1 Pro on 3D scenes, animations and websites in a video posted Oct. 23. He rated it above GPT-5.6 Sol and below GPT-6.1 Sol. "At least in these tests, I don't think it's quite there," he said of its standing against the frontier.

Moonshot's most concrete research demonstration began with a simulated human that collapsed within seconds. In an internal test, K3.1 Pro built a walking simulator around human-atlas, an open-source anatomical model, and iterated through repeated experiments. When reward adjustments failed to improve arm motion, the model investigated the simulator itself, found and corrected 36 errors in calculations governing shoulder-muscle joint rotation, then retrained the controller. The simulated human achieved alternating strides for 11.2 seconds. Moonshot ran this at maximum reasoning effort — temperature 1.0, top-p 0.95 — mostly on the Claude Code harness.

Artificial Analysis lists K3.1 Pro as proprietary until the weights ship, so cost-of-ownership claims for self-hosted deployments remain untestable. If the Nov. 4 weights release holds, research groups will have roughly two weeks of independent scrutiny before committing budget to either the API or their own GPUs.