October 25, 2026 | AI | By Mahesh

Moonshot AI released Kimi K3.1 Pro on October 23, 2026, a mixture-of-experts model with 2.8 trillion total and 104 billion active parameters, a 1 million token context window and API pricing of $2 per million input tokens and $10 per million output tokens [1][4]. Artificial Analysis scores it 55 on its Intelligence Index, up from 44 for Kimi K3 [3]. Moonshot promises open weights by November 4, 2026, and the weights and their license were not public at launch [2][4].

What Kimi K3.1 Pro is and what is new

Kimi K3.1 Pro is a post-trained version of Kimi K3. Moonshot calls it its most capable 3T-class open-weight model [1][4]. It is available in Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and accepts text, images and video through the API [1]. A Highspeed tier is pitched as 2 to 3 times faster at twice the price [2][3].

The headline gain is agentic coding. Moonshot reports 65.7 on Terminal-Bench 4.0, against 66.4 for Claude Opus 5.5, and 51.7 on SWE-Marathon, against 46.3 for Opus 5.5, with its own long-horizon coding benchmark rising from 55.95 for K3 to 61.2 [1]. It also reports a claim-level hallucination rate on its internal Kimi Hallu Bench Hard falling from 23.4% to 8.3%, against 5.5% for GPT-6 Astra [1].

Under the hood: experts, attention and serving

The model activates 16 of 896 experts per token using what Moonshot calls the Stable LatentMoE framework, and it keeps Kimi Delta Attention and Attention Residuals from K3 [1]. Thinking is always on, with reasoning effort set to low, high or max, and max is the default. Temperature is fixed at 1.0 and top-p at 0.95 [1][3].

The API is OpenAI-compatible with the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Output defaults to 131,072 tokens and can be raised to 1,048,576, which is a request limit and not proof that a full-window prompt and a full-window answer both work [1][3]. Moonshot advises returning the full assistant message with its reasoning history, because dropping it can destabilize output [1]. Public image URLs are rejected, so images go in as base64 or uploaded file IDs [1].

For the open-weight release, cyber training data was removed. On xdev-bench-hard, the shipped model solves 81.0% against 90.5% for an unfiltered checkpoint [1].

Benchmarks: independent index against vendor claims

Artificial Analysis gives Kimi K3.1 Pro 55 on its Intelligence Index, behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56 and ahead of Gemini 4 Argon and GPT-6 Astra at 53. The cost to run the index is $2.11 per task, against $5.98 for Opus 5.5 and $3.26 for Astra [3]. AI Weekly calls it second, and the table in Kingy's report that lists Sonnet 5.5 puts it third [2][3]. These are Artificial Analysis results and not Moonshot's.

Some results barely moved. SciCode went from 59.5 to 60.8, CritPt from 23.4 to 24.6, and AA-LCR, which tests about 100,000 tokens and not the full window, fell from 88.7 to 86.3 [3]. Gemini 4 Argon leads on AutomationBench-AA at 77.5 [3]. Artificial Analysis measured about 73.5 output tokens per second on the standard endpoint, and the Highspeed claim is unverified [3].

Moonshot's own table shows weaker spots. On Terminal-Bench Science 0.1 it scores 38.1 against 68.1 for GPT-6 Astra and 63.3 for Opus 5.5, and on BenchCAD with tools it scores 92.7 against 96.2 and 95.9 [1]. Several headline wins come from internal benchmarks, and the 65.7 on Terminal-Bench 4.0 used the Claude Code harness, while Artificial Analysis got 67.2 with mini-swe-agent [1][3].

A single hands-on test by Atomic Agent ran one small Python task three times per setup. Both Kimi K3.1 Pro and Opus 5.5 passed all 96 hidden tests. At low reasoning effort, Kimi cost $0.16 to $0.21 per task against $0.37 for Opus but took about 2.2 minutes against 69 seconds, and at max effort it cost $0.35 to $0.46 and took 4.9 minutes [4]. The author notes it was one small task on launch day.

Pricing, availability and how to access it

Opus 5.5 lists at $4 input and $20 output per million tokens, so K3.1 Pro is half that price [3][4]. Cache writes cost $2.00 for five minutes and $4.00 for one hour on the standard tier [3]. The API needs a minimum $1 top-up, and the model is on OpenRouter as moonshotai/kimi-k3.1-pro [1][4]. These are token rates and not task costs, and cache accounting differs between API interfaces [3].

Self-hosting is not possible until the weights appear. At 2.8 trillion parameters, 16-bit storage alone is at least about 5.6 terabytes, a calculated floor and not a published file size, and Moonshot recommended 64 or more accelerators for K3 [1][3]. The license has not been stated, and open weights do not automatically mean open source under the OSI definition [1][3].

Limitations, trade-offs and alternatives

Hallucination is still high on Artificial Analysis, at 43.9% against 53.2% for K3, and Moonshot's own lower figure uses an internal test [1][3]. Science-heavy terminal tasks trail GPT-6 Astra and Opus 5.5 by more than 25 points [1]. Kimi Code may fall back to K2.8 Preview if a model ID is mistyped, which could invalidate a comparison [3].

For other open-weight options, Depth Grid has covered Mistral Large 4, whose weights are due by the end of October, and for a pricing comparison with a closed model, Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks. Teams should rerun their own tasks at low reasoning effort before committing, because the cost savings in the one hands-on test came from lowering effort.

Sources

1. MarkTechPost. "Moonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context," October 23, 2026. marktechpost.com
2. AI Weekly. "Moonshot's Kimi K3.1 Pro scores 55 on AA, weights due Nov 4." aiweekly.co
3. Kingy AI. "Kimi K3.1 Pro: Specs, Benchmarks, Pricing and Comparisons," reporting Artificial Analysis results. kingy.ai
4. Atomic Agent. "Kimi K3.1 Pro vs Opus 5.5: Pricing, Benchmarks, Agent Test," October 23, 2026. atomicagent.io

Read More on Depth Grid

• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
• Nvidia in Talks to Deepen Stake in or Buy Reflection AI

Article by Mahesh | Depth Grid