Kimi K3.1 Pro Sets Terminal-Bench 4.0 Record at 67%, Lands 4th on Intelligence Index with 11-Point Gain
Moonshot AI released Kimi K3.1 Pro on October 23. It is the post-trained successor to Kimi K3 and the company’s new flagship, priced at $2/$10/$0.20 per million input/output/cache tokens — 33% below Kimi K3’s rates. Weights have not yet shipped, but Moonshot expects to release them shortly.
The headline result is in agentic performance, where the gains are large. On reasoning, coding, and science benchmarks, the improvements are real but moderate.
Intelligence Index and Agentic Benchmarks
On Artificial Analysis’s Intelligence Index, K3.1 Pro scores 55, placing 4th globally. The model is now 11 points above Kimi K3 (44) and trails only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56).
| Model | Intelligence Index | Terminal-Bench 4.0 | GDPval-AA ELO |
|---|---|---|---|
| Claude Opus 5.5 | 58 | 59.6% | 1,866 |
| Claude Sonnet 5.5 | 56 | 64.0% | — |
| Muse Wonder Full | 56 | — | 1,857 |
| Kimi K3.1 Pro | 55 | 67.0% | 1,846 |
| Kimi K3 | 44 | — | 1,534 |
Terminal-Bench 4.0 measures agentic task completion in a terminal environment. K3.1 Pro at 67% is the highest score recorded on the benchmark, 3 points above Claude Sonnet 5.5’s 64%. On GDPval-AA — Artificial Analysis’s broad agentic ELO — K3.1 Pro reaches 1,846, placing 3rd behind Claude Opus 5.5 and Muse Wonder Full, with a 312-point jump over Kimi K3. AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
Other Benchmark Gains
Gains outside agentic are present but smaller. Humanity’s Last Exam moves from 47% to 52%. CritPt rises from 23% to 25%. SciCode from 59% to 61%. The AA-Omniscience Index, which tracks hallucination rate and factual accuracy, rises from 20 to 30: hallucination rate drops from 53% to 44%, accuracy climbs from 48% to 51%. These are improvements across the board, but the agentic column is where K3.1 Pro separates from its predecessor.
Architecture and Pricing
The model is a 2.8-trillion-parameter MoE with 104 billion active parameters and a 1-million-token context window. That is the same parameter class as Kimi K3. Pricing is $2/M input, $10/M output, $0.20/M cache hit — 33% below K3 on all three rates. K3.1 Pro uses approximately 55,000 output tokens per Intelligence Index task, 14% more than K3’s 48,000, but the lower token prices result in a cost per task of $1.98, roughly equal to K3.
Open-Weight Significance
When weights are published, K3.1 Pro will be the leading open-weight model on the AA Intelligence Index. The current gap between the top proprietary model (Claude Opus 5.5 at 58) and the top open-weight model narrows from 12 points to 3 with K3.1 Pro in the picture. The only model within 3 points of K3.1 Pro above it on the Intelligence Index are three proprietary releases from Anthropic and Meta.
The Terminal-Bench 4.0 lead matters specifically for open-weight deployment: 67% on an agentic terminal benchmark, beating every proprietary model at that benchmark, is a concrete argument for self-hosted agentic infrastructure built on K3.1 Pro weights once they are available.