# Moonshot's Kimi K3.1 Pro comes within a point of Claude Opus 5.5 on Terminal-Bench 4.0, with open weights due by November 4

The 2.8-trillion-parameter mixture-of-experts model scores 65.7% against Opus 5.5's 66.4% and GPT-6 Astra's 58.2%. A separate harness study shows orchestration alone can raise a 27-billion-parameter open model above a closed frontier model on hard coding problems.

- Published: 2026-10-25T06:22:55.342Z
- Canonical: https://polylog.news/ai/2026-10-25/moonshot-s-kimi-k3-1-pro-comes-within-a-point-of-claude-opus
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Anthropic News](https://www.anthropic.com/claude-opus-5-5), [Hacker News](https://github.com/slee-persis/GVS5H)

Moonshot AI released [Kimi K3.1 Pro](https://www.marktechpost.com/2026/10/23/moonshot-ai-kimi-k3-1-pro-open-weight-model/) on October 23. It is a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters and a 1-million-token context window. The Beijing lab says it will publish the weights by November 4 ([Trending Topics](https://www.trendingtopics.eu/kimi-k3-1-pro-open-weight-model/)). The Russian-language channel [AI ML Big Data](https://t.me/ai_machinelearning_big_data/11200) reported that the model beats OpenAI's GPT-6 Astra on programming tests and is built for software development, agents and 3D modeling.

On agentic coding, the model performs close to the level of closed frontier models. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5, [The Next Web reported](https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber). Moonshot also reports 51.7 on SWE-Marathon v1.1 and 75.5 on DeepSWE v1.1. Artificial Analysis ranks the model fourth on its Intelligence Index with 55 points. By its calculation, the gap between the best open model and the best proprietary model would narrow from 12 points to 3 once the weights ship ([Trending Topics](https://www.trendingtopics.eu/kimi-k3-1-pro-open-weight-model/)).

Most of these figures come from the vendor, and no third party can run the model under its own harness until the weights are public. The knowledge-work comparison shows how much the harness and the benchmark version affect results. The leaderboard Moonshot cites puts K3.1 Pro at 1,846 Elo on GDPval-AA, behind Opus 5.5 at 1,866 and Meta's Muse Wonder Full at 1,857. Anthropic's own [launch page for Opus 5.5](https://www.anthropic.com/claude-opus-5-5) listed 1,846 on GDPval-AA v2.1 ([VentureBeat](https://venturebeat.com/technology/anthropic-releases-claude-opus-5-5-beating-fable-5-1-on-key-agentic-benchmarks-at-60-cheaper-api-price)). Anthropic released Opus 5.5 on September 22 and says it costs about 40% less to run than Opus 5 on typical workloads ([MindStudio](https://www.mindstudio.ai/blog/claude-opus-5-5-release)).

Moonshot disclosed one safety trade-off. Before the open release, the lab removed cyber-related data from midtraining and from the RL mix, giving up some vulnerability-research capability ([The Next Web](https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber)). K3.1 Pro builds on Kimi K3. Moonshot open-sourced K3's weights, its technical report and its training infrastructure on July 27 ([Beijing Daily](https://news.bjd.com.cn/2026/07/28/11889479.shtml)). Hong Kong media have reported that Moonshot plans to list in the city within the year ([Sing Tao](https://www.stheadline.com/zh-hans/realtime-china/3598323/Kimi-K3月之暗面开源28万亿参数AI模型-传计划年内赴港IPO)).

A separate result that reached Hacker News this week shows that orchestration can move coding scores as much as a new checkpoint. The [GVS5H repository](https://github.com/slee-persis/GVS5H) and its [arXiv preprint](https://arxiv.org/html/2608.26480v2) describe a training-free method. Fresh instances of one model split a problem and coordinate through a shared filesystem that holds a plan, notes and the current solution. On the 100 most recent hard LiveCodeBench problems, a locally served Qwen3.8-27B rose from 66.8% to 92.4% pass@1, compared with 90.4% for Anthropic's Claude Fable 5. The authors report gains of up to 25.6 percentage points across eleven models, and they also report that some models showed no gain or got worse ([TechBytes](https://techbytes.app/posts/gvs5h-five-qwen38-27b-models-match-claude-fable-5-on-livecodebench-hard/)).

## What this means

The mechanism is price and substitution. A downloadable model within one point of Opus 5.5 on an agentic coding benchmark lets inference hosts and enterprises run near-frontier coding work at the cost of their own hardware. That puts pressure on Anthropic's and OpenAI's per-token prices at the moment Anthropic is preparing a public listing priced on revenue growth. US controls on access to models do not reach weights posted from China, so governments and companies outside the US get a credible coding model regardless of US export policy. That supports the desk's thesis of two AI superpowers. The consensus treats each Chinese open release as a capability event. The GVS5H data suggests the harness now accounts for a large share of measured coding performance, which makes leaderboard gaps less reliable in both directions. The nearest precedent is DeepSeek-R1 in January 2025: Nvidia lost about 17% in one session, and OpenAI released the cheaper o3-mini within a week. The desk's read is that US inference providers will likely host K3.1 Pro within weeks of the weights appearing, and at least one closed lab will cut the price of its coding tier before year-end. This depends on Moonshot releasing the full weights under a license that allows commercial hosting. Independent Terminal-Bench 4.0 runs that land well below 65.7% would prove this read wrong.

## What to watch

- Whether Moonshot posts the full K3.1 Pro weights by November 4, and under what license. A restrictive license would keep US and European inference providers from hosting it.
- Third-party Terminal-Bench 4.0 and SWE-style runs once the weights are public. They will show whether the 65.7% holds outside Moonshot's own harness.
- Price changes on Anthropic's and OpenAI's coding-focused tiers in the weeks after the weights ship. A cut would show that open releases now cap what closed labs can charge.