[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed

What it claims

Moonshot's Kimi K3.1 Pro is a post-trained version of Kimi K3 that uses much more RL compute. It has the same base model, with joint RL across "hundreds of downstream domains." Moonshot advertises full-project coding in Unreal, Unity and Godot, parametric Blender and CAD work, financial modelling with "65% fewer hallucinations," and long-running agents. The finance figure has no stated baseline or protocol. The launch also added Kimi Hosted Agents, a managed harness with a 24/7 execution environment for long-running and multi-agent work. No Flash model was announced.

Evidence

Artificial Analysis (AA) reports these gains over K3:

• Hallucination: on AA-Omniscience, accuracy rose from 48% to 51% and hallucination rate fell from 53% to 44%.
• Open versus proprietary gap: AA says that once weights ship, the gap on its Intelligence Index between the best open model and the best proprietary model narrows from 12 points to 3.
• Cyber: the Cyber Index score is 64%, up 23 points from K3's 41%. Most of the gain came from CyberGym, which rose from 58% to 95%. This is the highest publicly accessible result. Trusted-access GPT-6 Sol Daybreak Blue scored 69%. AA saw no safety refusals on its tasks.

Moonshot says it removed cyber-related data from mid-training and RL. Posts quoting its report give an internal cyber score drop from 90.5% to 81% relative to the counterfactual model. Coding and other capabilities still improved. The data removal therefore lowered cyber capability against Moonshot's own baseline but did not produce API refusals.

Arena results:

• WebDev: 1,692 points, against K3's 1,655 at #17.
• Other categories: Simulations rose from #20 to #9, and Brand & Marketing from #9 to #7.
• Peer comparison: Arena says WebDev performance roughly matches Opus 5 Max, not Opus 5.5.

Results outside conventional coding are weaker:

• AA-3D: early preferences place K3.1 Pro below its coding standing. Voting was still open.
• RuneScape Bench: Max Bittker reports costs comparable to Argon and performance close to DeepSeek V4.1 Flash, about Opus 5.0 level.
• Compiler work: one user (@QuixiAI) found it behind Fable 5.1 and Opus 5.5.

How to use it

The model is available now through kimi.com, the Kimi app, Kimi Work, Kimi Code and the API, in Standard and Highspeed versions.

Standard pricing:

• Per-token rates: $2 per million input tokens, $10 output and $0.20 cached input. K3 cost $3, $15 and $0.30.
• RMB rates: 13 uncached input, 68 output and 1.3 cached input per million tokens.
• Highspeed: claims 2 to 3 times faster output at the same intelligence, for twice the Standard price.

Per-task cost does not fall as much as per-token rates:

• Token use: on the Intelligence Index, the model used 55K output tokens per task against K3's 48K. On AA's cyber suite it used about twice K3's input tokens, which raised task cost despite the lower rates.
• Arena's measure: median task cost is $1.23. That is 74% below Opus 5 Max, 24% below GPT-6 Sol Max and 47% below Qwen3.8-Max.
• Terminal-Bench: cost claims conflict. One poster says about one-third of Opus 5.5's cost; Cline says about half at near parity.

Cline offers access through ClinePass at $9.99 per month.

Self-hosting needs to hold the full 2.8T expert pool, not just the 104B active parameters. The newsletter estimates about 5.6 TB at 16-bit or 1.4 TB at 4-bit for the weights alone, before serving overhead.

Limits

Weights were unpublished at launch and are promised for 4 November 2026. The license, commercial-use terms, checkpoint precision and inference requirements are not stated. The page also gives no layer layout, routing configuration, attention design or training compute.

Moonshot's own benchmark table is not an independent replication. Commenters raised three issues with it:

• It makes heavy use of the Claude Code harness.
• Its DeepSeek Terminal-Bench numbers appear to come from mixed sources.
• Its Terminal-Bench figure of 65.7% differs from AA's rounded 67%.

@teortaxesTex estimates the model may be 30 or more times costlier to serve than DeepSeek Flash. This is an estimate, not a measurement. Serving quality also varies by provider: Vals found K3 scored 60.6% on legal tasks via Baseten against 75.8% via Moonshot, and Moonshot publishes an endpoint verifier. Separately, there are unverified allegations that launch promoters were told to omit paid-partnership labels. No Moonshot response appears in the issue.