Trending AI Insights: Did Kimi K3.1 Pro Just Cut the Gap to 3 Points?

On 23 October, Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series, alongside a Highspeed variant that runs roughly two to three times faster. It landed fourth on the Artificial Analysis Intelligence Index with a score of 55, behind only Anthropic's Claude Opus 5.5, Sonnet 5.5 and Muse Wonder Full — three closed models.

The more consequential number is a different one: the intelligence gap between the best open-weight model and the best proprietary one narrowed from 12 points to 3. Once the weights ship on 4 November, Artificial Analysis expects K3.1 Pro to become the highest-scoring open-weight model ever released.

Specs and scores

• Scale: 3-trillion-parameter class, 2.8 trillion total, a mixture-of-experts layout activating about 104 billion parameters per token, with a 1-million-token context window.
• Terminal work: 65.7% on Terminal-Bench 4.0, above GPT-6 Astra at 58.2% and 0.7 points behind Claude Opus 5.5 at 66.4%. That benchmark tests whether a coding agent can actually finish jobs inside a real terminal.
• Long-horizon coding: 61.2% on the in-house Kimi Code Bench 3.0, up from 55.95% for K3.
• 3D and spatial reasoning: 92.7% on BenchCAD with tools (Opus 5.5 scores 96.2%, GPT-6 Astra 95.9%). A companion feature, 3D Make, previews, edits and exports models for printing.
• Finance: 64.4 on OfficeQA Pro v2, which Moonshot calls state of the art. On three internal finance benchmarks, modelling, research and business operations improved 14%, 20% and roughly 4.2x over K3.
• Hallucination: on the internal Kimi Hallu Bench Hard stress test, the single-statement error rate fell from 23.4% to 8.3%.

How it was built

K3.1 Pro is not a new base model. It is a re-post-trained K3: far more post-training compute, plus large-scale joint reinforcement learning across hundreds of downstream domains. Moonshot's emphasis is less on accelerator count than on tuning the whole stack — RL framework, long-context inference, training efficiency, sandbox infrastructure — so each increment of RL compute returns a stable, predictable gain. That mirrors the industry consensus since 2024: pretraining scale is hitting diminishing returns, and post-training plus RL is where the marginal dollar still buys intelligence.

Half the price, but not half the bill

API pricing is $2 per million input tokens and $10 per million output, about half of Opus 5.5's rate card. In China the list is 13 yuan per million input, 1.3 yuan when the cache hits, and 68 yuan output; Highspeed doubles it.

There is a catch that headline comparisons tend to swallow. Half price per token does not mean half price per finished task. Artificial Analysis measures K3.1 Pro at roughly $1.98 per Intelligence Index task versus $2.00 for K3 — essentially level — because it emits about 55,000 output tokens per task, 14% more than its predecessor. OrcaRouter separately notes that the Terminal-Bench numbers were produced at different reasoning settings, so cross-model comparisons deserve caution.

There is also a tradeoff that gets less attention. Before publishing weights, Moonshot stripped cyber-related data out of mid-training and RL. The cost is measurable: on its internal xdev-bench-hard test for vulnerability discovery and exploitation, the final model scored 81.0%, down from 90.5% at the earlier checkpoint and below frontier proprietary models that have been publicly available for at least three months. The stated principle is that open weights should not push the ceiling of offensive cyber capability higher, while defensive usefulness should survive.

The signal is a curve, not a single model

Zoom out and the trajectory is steep. Moonshot was founded in 2023, started as a long-context chatbot, and reached a $50 billion valuation in under four years. Hugging Face's 2026 report found that open models developed in China accounted for 41% of platform downloads, ahead of the United States at roughly 35-36.5%. Axios put it bluntly: America's commanding lead in advanced AI is gone.

The capital side is accelerating in parallel. On 22 October, DeepSeek closed a round of about 100 billion yuan (roughly $15 billion) at a pre-money valuation near 500 billion yuan, with CATL and Tencent among the largest investors. Its annualised revenue run rate has more than doubled from under $500 million to above $1 billion, and it has retained CITIC Securities to prepare a STAR Market listing. The contest between the two countries has moved from whose benchmark is higher to who can keep buying compute, keep developers, and turn models into cash flow.

For developers, the day open weights actually catch closed ones is the day the price war starts in earnest. Terminal coding agents are already the must-win lane: K3.1 Pro trails by 0.7 points and costs half as much. If you want to find the tools that genuinely run inside your terminal and your workflow — coding agents, code review, automated testing, local deployment stacks — start at aibookmarks.net, one of the most comprehensive AI tool directories, which sorts global AI tools by the problem you need solved rather than by launch date.

Three things to watch

• The 4 November weight release. Licence terms, quantised builds and the fine-tuning ecosystem decide whether this becomes a DeepSeek-style open-weight moment.
• Highspeed in practice. Whether the 2-3x speed comes out of latency or out of quality will only show up in hands-on testing.
• Who answers next in terminal coding. The bar now sits at 65.7%.

The gap is now three points. The question is no longer whether open weights catch up, but how fast the price of that gap collapses.

Original author: CGTN Tech Desk, drawing on Moonshot AI's release notes and Artificial Analysis evaluations.