By Hamza Faisal

Moonshot AI has unveiled Kimi K3.1 Pro, the first major upgrade since Kimi K3 gave the Beijing startup its "DeepSeek moment" over the summer. The new model went live on October 23 across Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the company's API — and its sights are clearly set on the closed frontier models from OpenAI and Anthropic.

K3.1 Pro keeps K3's architecture: 2.8 trillion total parameters with 104 billion active per token in a mixture-of-experts design, text and image input, and a 1-million-token context window. The upgrade is all in the training. Moonshot says it ran substantially more post-training compute and reinforcement learning across "hundreds of downstream domains," tuned for long-horizon agentic coding, 3D creation and financial research.

Benchmarks: near-parity on coding, a gap on science

Moonshot's self-reported numbers put K3.1 Pro within striking distance of the best closed models on coding agents. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, K3.1 Pro scored 65.7% — ahead of OpenAI's GPT-6 Astra (58.2%) and just behind Anthropic's Claude Opus 5.5 (66.4%). It topped SWE-Marathon outright at 51.7, ahead of Opus 5.5's 46.3, and improved from 55.95 to 61.2 on the company's internal long-horizon coding benchmark.

The picture is less flattering on science: on Terminal-Bench Science 0.1, K3.1 Pro scored 38.1 against 68.1 for GPT-6 Astra and 63.3 for Opus 5.5. Most scores were also run at maximum reasoning effort inside Anthropic's Claude Code harness, so independent verification will matter.

Finance, 3D modeling, and fewer hallucinations

Moonshot is pitching K3.1 Pro hardest at two commercial niches: finance and 3D. It claims a state-of-the-art 64.4 on OfficeQA Pro v2, a 7.2-point gain on financial modeling and 11.6 points on financial research over K3, and roughly 4.2x better simulated business operations. Demo outputs include a fully linked NVDA DCF/LBO model with traceable formulas, an investment-committee deck for Chinese battery giant CATL, and a quantum-computing sector report with source-level citations.

On 3D, the model scored 92.7% on BenchCAD with tool use — close to Opus 5.5's 96.2 and GPT-6 Astra's 95.9. Moonshot showed it generating game environments in Unreal, Unity and Godot, scripting Blender and CAD tools for parametric 3D modeling, and fixing 36 errors in an anatomically detailed walking simulator until its virtual character could walk for 11.2 seconds.

Hallucinations, the eternal enterprise objection, are down sharply: claim-level errors fell from 23.4% to 8.3% on the company's internal hard test — a roughly 65% drop, though GPT-6 Astra still scores a lower 5.5% on the same test.

Cheaper than K3, open weights in November

API pricing undercuts K3: $2 per million input tokens and $10 per million output tokens (down from $3/$15), with cached input at $0.20. A companion Kimi K3.1 Pro Highspeed variant promises 2–3x faster responses at $4/$20. Moonshot has pledged to release the full model weights by November 4 — a move that keeps pressure on the open-weight race.

One deliberate choice stands out: Moonshot stripped hacking-related data from mid-training and reinforcement learning before the open-weight release. The shipped model scores 81.0% on the company's internal offensive-cyber bench versus 90.5% for an unfiltered checkpoint — a self-imposed ceiling that shows frontier labs are now pre-empting the safety battles themselves.

By Hamza Faisal