ModelBrief · Kimi K3.1 Pro · Topic Page
Last updated Nov 20, 2026

Kimi K3.1 Pro Benchmarks, All of Them: AA 55/46 and TB 4.0 at 67.2%

ModelBrief · Published 2026-11-16 · Fact-check 2026-11-16 · Reference: Artificial Analysis, LMArena, LiveBench, Agent Arena, Moonshot tech blog

K3.1 Pro's benchmark evidence has two layers: the independent boards are the thickest layer of any topic on this site — Artificial Analysis, LMArena, LiveBench and Agent Arena corroborate each other; the vendor blog's 14 benchmarks are the almost-entirely-self-reported layer. This article covers the independent numbers first, then the vendor table, then the two figures most easily written wrong: AA's two index versions, and the weakened cyber score.

The AA Index: 55 and 46 are both real, and cannot be compared

The Artificial Analysis Intelligence Index is this model's most important independent evidence, but it has two versions that must always be stated together: v4.3.2 (Oct 23, 2026) scored 55, No. 4 of 215 models, at $2.11 per task; Opus 5.5 scored 58, Sonnet 5.5 56 and GPT-6 Astra 53 on the same version. Around November 13, AA re-versioned the index to v5.0 (a new evaluation set) and the same model scored 46 — still the best open-weights model on the board. Both numbers are real; the "drop" from 55 to 46 is a re-version, not a weaker model. Any citation that writes one number without its version is wrong — readers will find the other number elsewhere and assume one of them is fake.

AA Intelligence Index v4.3.2 comparison (Oct 23, 2026) Kimi K3.1 Pro Opus 5.5 Sonnet 5.5 GPT-6 Astra Opus 5.5 58 Sonnet 5.5 56 Kimi K3.1 Pro 55 GPT-6 Astra 53 AA Intelligence Index v4.3.2, Oct 23, 2026; orange is Kimi K3.1 Pro On the v5.0 re-version K3.1 Pro scores 46, still best open weights — the versions cannot be compared directly
AA Intelligence Index v4.3.2 (points; Artificial Analysis, Oct 23, 2026; orange is Kimi K3.1 Pro)

AA's other boards: terminal, cyber and coding agents

AA's other boards are also independent measurements: Terminal-Bench 4.0, run by AA itself, scored 67.2%, the highest of every model on its page — while the vendor blog reports 65.7% for the same benchmark, so the vendor under-reported this time (though the two runs may use different settings, and should not be subtracted as if they were the same harness). The Cyber Index (Nov 13) gives 64, against Hy4's 58 and DeepSeek-V4.1-Flash-1017's 57; CyberGym end-to-end gives 94.7%, No. 1 of 26, ahead of GPT-6 Sol at 93.1%. The Coding Agent Index v2.1 (Nov 13) gives 49 (Kimi Code CLI), behind Opus 5.5 and Fable 5.2 at 58, Carbon at 55 and Astra at 53 — on coding agents, K3.1 Pro clearly trails the closed flagships.

Terminal-Bench 4.0 comparison (vendor-reported vs AA-measured) K3.1 Pro (AA-measured) K3.1 Pro (vendor) Rivals (vendor table) K3.1 Pro (AA-measured) 67.2% Opus 5.5 (vendor table) 66.4% K3.1 Pro (vendor) 65.7% GPT-6 Astra (vendor table) 58.2%
Terminal-Bench 4.0 (%; AA-measured 67.2% vs vendor-reported 65.7% may use different settings — do not subtract directly; rival scores from the vendor table)

LMArena, LiveBench and Agent Arena: three boards, one picture

LMArena Text: 1488.0, No. 24 (style control, 3,995 votes, preliminary, ±13.0); WebDev 1687 Elo (HTML 1689, No. 10 of 104). LiveBench: coding 67.8% ($1.395 per task, $2.057 per successful task), reasoning 91.1% (No. 8 of 31), browser 75.9% (No. 13). Agent Arena (arena.ai): No. 1 open weights, No. 13 overall, net improvement +5.1% (Max), on 17,500+ real long-horizon agent sessions; Bash Reliability +7.9%, against K3's −4.5%. DuelLab GameBench 2: 78.3, No. 10, Max Intensive 84.0. The four boards tell a consistent story: near-closed-flagship overall, with agents and reasoning as strengths; but on LMArena it leads Kimi K3 by only 0.6 points (1488.0 vs 1487.4, and 1.2 behind Muse Spark at 1489.2) — inside the confidence interval, not significant. On general chat, post-training opened almost no gap.

The vendor's 14 benchmarks: finance and hallucination shine, BenchCAD lags

The official blog's table (max reasoning, temp 1.0, top-p 0.95) carries 14 tests, almost all vendor-reported: Terminal-Bench 4.0 65.7% (Opus 5.5 66.4%, GPT-6 Astra 58.2%); BenchCAD with tools 92.7% (Opus 5.5 96.2%, GPT-6 Astra 95.9%, Fable 5.1 92.6%), without tools 72.8% (Opus 5.5 73.0%); Kimi Code Bench 3.0 long-horizon (internal) 61.2% vs K3's 55.95%; OfficeQA Pro v2 (finance) 64.4, SOTA per Moonshot; Frontier Finance 58.6; internal finance benchmarks vs K3: Modeling 57.6 vs 50.4, Research 70.4 vs 58.8, Business Operation 37 vs 8.9 (~4.2×); Kimi Hallu Bench Hard claim-level hallucination: with tools 8.3% (K3 23.4%), without tools 5.4% (GPT-6 Astra 4.2%, Opus 5.5 6.8%). Read the table in two layers: the finance and hallucination gains are large (Business Operation 8.9 → 37, hallucination 23.4% → 8.3%) — real post-training returns; BenchCAD and TB 4.0 trail Opus 5.5, so it is not yet an all-round flagship.

The cyber weakening: 90.5% to 81.0%, a choice the vendor stated itself

xdev-bench-hard pass@3 (cyber): the uncleaned checkpoint 90.5%, the released model 81.0%; K3 33.3%, DeepSeek-V4.1-Flash 85.7%. The blog explains why: Moonshot removed cyber-related data from mid-training and RL so the open weights would not extend the "absolute frontier of cyber capability." The vendor's own framing: the released model's cyber capability "remains behind that of frontier proprietary models that have been publicly available for at least three months." This is a deliberate downgrade, not a capability ceiling — but the released 81.0% is still below DeepSeek-V4.1-Flash's 85.7%, so on cyber tasks it is not the open leader. Meanwhile AA's CyberGym end-to-end run gives 94.7%, No. 1 of 26 — a different harness (AA's end-to-end run vs the vendor's pass@3), not directly comparable, and neither invalidates the other.

The weak spots: slow, and trailing on coding agents

The weak spots, collected: output speed is slow — AA's two measurements conflict (~49.7 vs ~73.5 tok/s, around No. 186 of 192 on the speed board), and eesel.ai calls it "smart and cheap per token, but slow"; the Coding Agent Index is 49, behind Opus 5.5 and Fable 5.2 at 58 and Carbon at 55; BenchCAD with tools is 92.7%, behind Opus 5.5's 96.2%; the Highspeed "2–3× faster" claim has no published tok/s figure and cannot be verified; the LMArena gap over K3 is not significant. These weak spots don't invalidate the 46-point "best open weights" claim, but they set the boundaries: agent and finance tasks are worth trying; speed-sensitive long-output tasks need a test first.

FAQ

Is the AA Index 55 or 46?

Both are real: 55 is v4.3.2 (Oct 23, 2026), 46 is v5.0 (~Nov 13, re-versioned with a new evaluation set). Scores from two index versions cannot be compared directly; the v5.0 score of 46 is still the best open-weights model on the board.

Is Terminal-Bench 4.0 65.7% or 67.2%?

65.7% is the vendor blog's self-reported figure; 67.2% is Artificial Analysis's own measurement of the same benchmark, and the highest on its page. The two runs may use different settings, so they should not be subtracted directly; the vendor under-reported this time.

Does the 81% cyber score mean the model is weak?

Not a capability ceiling — a deliberate weakening: Moonshot removed cyber-related training data, and the uncleaned checkpoint scored 90.5%. But the released 81.0% is still below DeepSeek-V4.1-Flash's 85.7%, so on cyber tasks it is not the open leader.

How fast is its output, exactly?

There is no credible single figure: AA's two measurements are about 49.7 and about 73.5 tok/s, in conflict, with a speed-board rank around 186 of 192. The Highspeed "2–3× faster" claim has no published tok/s figure. This site adopts no single speed number.

ModelBrief's take: K3.1 Pro's benchmark profile is remarkably consistent: near-closed-flagship overall, agents and finance as strengths, speed and coding agents as weak spots, and the thickest evidence base on the site. The key to reading its scores isn't memorizing 55 or 46 — it's remembering the version number — and remembering that 81.0% is the figure after the vendor's deliberate downgrade, while 90.5% is the uncleaned checkpoint's real capability.