By ByteBot
Moonshot AI released Kimi K3.1 Pro yesterday, and the headline number is this: it jumped 11 points on the Artificial Analysis Intelligence Index over Kimi K3 in roughly three months. That puts it at 55 — tied with Claude Sonnet 5.5 and 3 points behind Claude Opus 5.5. On November 4, the full weights drop. When they do, it will be the highest-ranked open-weight model you can actually run.
What the Benchmarks Actually Show
The vendor numbers look good, but the independent ones are more useful. Artificial Analysis puts K3.1 Pro at 55 on its Intelligence Index — fourth globally behind Opus 5.5 (58), Sonnet 5.5 (56), and one other proprietary model. On Terminal-Bench 4.0, Moonshot reports 65.7%, ahead of GPT-6 Astra at 58.2% and essentially tied with Opus 5.5 at 66.4%.
3D modeling is the surprise. On BenchCAD — a benchmark that tests 3D reconstruction via CAD code generation — K3.1 Pro scored 92.7%. That is within 3 points of Opus 5.5 (96.2%) and GPT-6 Astra (95.9%), which means Moonshot is now frontier-competitive in a domain most people did not expect. The model integrates directly with Blender, Unreal Engine, Unity, and Godot.
Hallucination rates improved significantly too. On Moonshot’s internal stress-test benchmark, K3 came in at 23.4% claim-level errors. K3.1 Pro is at 8.3%. That is the kind of improvement that matters for production pipelines where wrong answers are expensive.
The Deliberate Handicap
Here is the part worth paying attention to. Moonshot deliberately degraded K3.1 Pro’s cybersecurity capability. They identified and removed cyber-related training data from both training stages. Their internal unrestricted checkpoint scores 90.5% on xdev-bench-hard, an adversarial exploit benchmark. The released model scores 81.0% — a 9.5-point intentional reduction.
The stated rationale: open weights enable far more teams to audit code, which is a net security positive. Giving the open-weight community an extreme offensive cyber tool is not. That is a defensible position, and it is a direct counter to the argument that open-weight models are inherently dangerous because anyone can fine-tune out safety layers.
The somewhat ironic result: despite the deliberate capability reduction, K3.1 Pro still scored 64% on Artificial Analysis’ Cyber Index — the highest publicly accessible result on that benchmark. Removing training data lowered the ceiling but did not eliminate the capability.
The Cost Reality
K3.1 Pro costs $2.00 per million input tokens and $10.00 per million output tokens. DeepSeek V4.1 Flash costs $0.15 and $0.60 respectively — 13x cheaper on input. The Intelligence Index gap between them is 16 points (55 vs. 39), but for long agentic sessions, DeepSeek actually beats Kimi — 77% versus 71% — because Kimi spends more tokens thinking per turn.
The routing decision is not complicated: send classification, extraction, summarization, and first-pass drafts to DeepSeek. Send multi-step coding tasks, agent orchestration, and work requiring error recovery across turns to Kimi K3.1 Pro. Both offer 1M context windows and multimodal input, so the choice reduces to task complexity and budget. Cached input drops to $0.20 per million — a 90% discount — which helps significantly for repeated large-document workflows.
How to Access It Now
The API is live today via OpenRouter and directly through Moonshot. The model identifier is kimi-k3.1-pro via the standard API (note: Kimi Code uses k3.1-pro without the prefix). Reasoning effort is always on — default is max, but you can set it to low or high depending on task requirements. K3.1 Pro Highspeed delivers 2-3x faster output but consumes twice the quota.
The full model weights arrive November 4. At 4-bit precision that is approximately 1.4TB of storage before serving overhead — not a consumer-device model. License terms have not been announced yet. Watch that carefully before building deployment plans around it. The official Moonshot announcement has full technical details, and the K3.1 Pro vs. DeepSeek cost breakdown is worth reading before making routing decisions.
The Bigger Picture
The open-weight gap to frontier closed models was approximately 12 Intelligence Index points before this release. K3.1 Pro narrows it to roughly 3. That is not nothing — 3 points on this scale represents meaningful capability differences on hard tasks — but it signals that the gap between what you can self-host and what you can only access via API is approaching a practical threshold. Ten days from now, developers will be able to run what is effectively near-frontier intelligence on their own infrastructure. That is a materially different world than six months ago.
