As of 24 October 2026Kimi K3.1 Pro launched 23 Oct 2026Weights due by 4 Nov 2026
Moonshot’s Kimi K3.1 Pro is the strongest Chinese model yet on public tests. On the main independent composite it sits 3 points behind Claude Opus 5.5 (55 vs 58) — but that single number hides a split: near-parity on long, tool-driven agent work, and a clear deficit on unaided knowledge and strict-completion coding. This page lets you see both, then set your own weights.
There is no single “gap”. Read it three ways — composite points, task profile, and calendar time — and you get three different answers. The honest summary is: months, not generations; and on agentic work, effectively now.
How to use this page: §§1–6 show the evidence. The estimator at the end converts your own task mix into a gap in points and months. Nothing here uses private evals — sources are linked under every chart.
The Artificial Analysis Intelligence Index is the cleanest single-number comparison available: one independent harness, many evals (reasoning, knowledge, coding, agents), run at each model’s max reasoning. K3.1 Pro jumped +11 points over Kimi K3 (44 → 55) in one post-training step.
Higher is better. Red = Chinese labs, navy = Western labs. Only three models sit above Kimi K3.1 Pro — all Western, all closed. Two Western flagships (GPT-6 Astra, Gemini 4 Argon) sit below it.
Source: Artificial Analysis, Kimi K3.1 Pro launch benchmarks (23 Oct 2026). Precise v4.3.2 scores where published: Opus 5.5 57.6, Sonnet 5.5 56.0, Fable 5.1 53.4, Astra 52.7, Argon 52.6. Opus/Sonnet run at “max with fallback”, others at “max” / “high” — conditions not perfectly uniform. AA: once weights ship, K3.1 Pro will be the top open-weight model and the open-vs-proprietary gap narrows 12 → 3 points.
Reading the tail matters too. Before K3.1, the best Chinese scores clustered at 44–46 (Kimi K3 44, GLM-5.3 45, Qwen3.8 Max 45, MiMo-V2.6-Pro 46) against 58 at the top — a 12–14 point chasm. K3.1 Pro did not nudge that cluster; it left it. The question is whether the jump is broad capability or a narrow agentic surge. Answer: mostly the latter (§4).
Moonshot’s launch table compares K3.1 Pro against five peers across coding, knowledge-work, and science tests. Sorted by Kimi minus best Western score, the shape is unmistakable: long-horizon agent tasks at the top, strict-completion and unaided-knowledge tasks at the bottom.
Right of zero = Kimi leads the best Western model on that test. Left = trails. “Best Western” = max of Opus 5.5 / Fable 5.1 / Astra per row, so every bar is Kimi against the strongest rival, not the average. Hallucination inverted (lower is better, shown as advantage).
Source: Moonshot K3.1 Pro tech blog (23 Oct 2026), as tabulated by XenoSpectrum and explainx. Kimi run mostly in the Claude Code harness at max effort; competitor scores are compiled maxima from vendor blogs / official leaderboards, not rerun in one harness. Treat ±2 pts as a tie. Full numbers in the table below.
| Benchmark (Moonshot table) | Kimi K3.1 Pro 🇨🇳 | Opus 5.5 | Fable 5.1 | GPT-6 Astra | DeepSeek V4.1 🇨🇳 | GLM-5.3 🇨🇳 | Gap vs best West |
|---|---|---|---|---|---|---|---|
| SWE-Marathon v1.1 — long-horizon coding | 51.7 | 46.3 | 45.6 | 42.5 | 40.0 | 42.5 | +5.4 |
| HLE-Full with tools — hard knowledge + search/code | 70.0 | 67.7 | 65.6 | 57.2 | — | — | +2.3 |
| OfficeQA Pro v2 — financial PDFs as images | 64.4 | 62.2 | 54.5 | 54.4 | 40.4 | 31.1 | +2.2 |
| DeepSWE v1.1 — deep repo coding | 75.5 | 74.2 | 67.4 | 74.1 | 77.0 | 66.9 | +1.3 |
| Terminal-Bench 4.0 — terminal agent tasks | 65.7 | 66.4 | 57.9 | 58.2 | 42.3 | 41.8 | −0.7 |
| BenchCAD with tools — 3D via CAD code | 92.7 | 96.2 | 92.6 | 95.9 | 81.6 | — | −3.5 |
| BenchCAD without tools — single-shot | 72.8 | 73.0 | 60.6 | — | 41.2 | — | −0.2 |
| Hallu Bench Hard, with tools ↓ lower better | 8.3 | 15.0 | 23.3 | 5.5 | 23.3 | 22.6 | −2.8 |
| HLE-Full without tools — closed-book | 51.1 | 59.8 | 57.6 | 53.6 | — | — | −8.7 |
| ProgramBench binary (≥95% = pass) — strict finish | 31.5 | 65.0 | 53.5 | 50.0 | 29.1 | 19.0 | −33.5 |
Pink = best Chinese score in row · Blue = best Western score. “—” = not reported in Moonshot’s table. Internal-only rows (Kimi Code Bench 3.0 Long-horizon 61.2% vs K3 55.95%; Hallu Bench is also internal) are excluded from the chart except where cross-model numbers were published. Note DeepSeek beats Kimi on DeepSWE — “China” is not one lab.
The single most diagnostic result in the launch table is Humanity’s Last Exam run both ways. Give every model search and code execution and Kimi wins. Take tools away and it finishes last of the four. That is not a rounding error — it is an 18.9-point swing.
Each pair is the same exam. The gap between the two bars is not a pure “tool boost” (conditions differ), but the rank flip is real in Moonshot’s own numbers.
Source: Moonshot table via XenoSpectrum (23 Oct 2026). Max reasoning; “max with fallback” for non-public rivals. Do not compare with AA’s HLE 52% (§4) — different harness/subset.
3D reconstruction via CAD code (1,000-case subset, Voxel IoU). Without tools Kimi essentially ties Opus (72.8 vs 73.0); with tools both jump ~20 pts and Opus pulls 3.5 pts ahead.
Source: Moonshot tech blog. With-tools via Claude Code harness; without-tools single-shot. GPT-6 Astra reports no without-tools score. A high BenchCAD score ≠ manufacturing-ready drawings.
Artificial Analysis ran the same evals on K3 and K3.1 Pro, so the deltas are comparable. Agentic scores exploded; knowledge and science scores barely moved. Moonshot says K3.1 Pro is K3 with “substantially larger post-training compute” — joint RL across hundreds of domains — and the fingerprint matches.
Light = K3 (July 2026) · Solid = K3.1 Pro (Oct 2026). GDPval-AA is Elo (right-hand annotation), everything else is %. Terminal-Bench went from worst-in-class to best-in-class in one step.
Source: Artificial Analysis (23 Oct 2026). GDPval-AA: 1,534 → 1,846 Elo (+312), 3rd behind Opus 5.5 (1,866) and Muse Wonder Full (1,857) — not plotted on the % axis. AA-Omniscience Index 20 → 30; hallucination rate 53% → 44%; accuracy 48% → 51%. Moonshot’s internal Hallu Bench Hard (different test): 23.4% → 8.3% claim-level hallucination.
Why this matters for the gap debate. A +54-point Terminal-Bench jump (13% → 67%) in ~3 months tells you the old 44 was understating the base model’s latent agent ability — or that RL + harness fit can move agent benchmarks extremely fast. Either way, extrapolating “China is N months behind” from a pre-K3.1 snapshot is now stale. It also warns against over-trusting any one benchmark: scores that can quintuple in a quarter are measuring trainable, harness-sensitive skills, not fixed national capability.
The cyber footnote. Moonshot deliberately removed cyber-related data from mid-training and RL for the open-weight release. On its internal xdev-bench-hard (pass@3), the uncapped checkpoint scores 90.5% vs 81.0% shipped — i.e. the released model is intentionally below what the same pipeline can produce, and Moonshot states it remains behind frontier proprietary models public for ≥3 months. The “gap” on offensive cyber is therefore partly a policy choice, not a capability ceiling. (Moonshot blog, Cybersecurity §.)
“Months behind” is computed by asking: when did the US frontier first reach the capability China has today? Different groups, indexes, and end-dates give 3–8 months. All of them predate K3.1 Pro except the last row — which is why they now read as upper bounds for the composite.
Bars show point estimates; whiskers show published ranges where given. Methods differ (Epoch ECI trend fit, NIST evals, AA index matching) — compare the spread, not just the centres. The bottom bar is this page’s K3.1-updated composite (see method note).
Sources: Epoch AI Capabilities Index — Chinese models trail US frontier by 7 mo avg since 2023 (range 4–14); open-vs-closed ~4 mo in 2026, 8 ECI pts ≈ GPT-5→GPT-5.5 (summary, Oct 2026). IFP trend fit (US 1.31 vs China 1.18 ECI/mo, thru K3 Jul 2026 / Astra Sep 2026): ~6.5 mo (IFP). NIST CAISI (May 2026): DeepSeek V4 Pro ~8 mo behind. Mozilla State of Open Source AI (Sep 2026): best Chinese open ~4.4 mo behind US closed. Reuters via officials: Kimi K3 3–6 mo. Stanford AI Index 2026 (different units): top-model gap ~2.7–3%, down from 15% early 2026 — consistent with AA’s 3/58 ≈ 5%.
Two cautions on “months”. First, cadence asymmetry: US frontier labs ship every 2–4 months; DeepSeek took 16 months from V3 (Dec 2024) to V4 (Apr 2026). A 3-month capability lag with a slower shipping cadence is more fragile than the number suggests — the frontier moves while you catch up. Second, the compute backdrop: these lags persist despite a large US compute advantage and export controls, which is precisely why analysts treat the small lag as strategically significant. Benchmarks alone cannot tell you whether the lag is stable, closing, or about to widen — only the next two release cycles will.
List price per token flatters verbose models. What matters is dollars per completed task. On Artificial Analysis’s task suite, K3.1 Pro costs about a third of Opus 5.5 per task for 95% of its Index score — and its weights will be downloadable.
Each dot is a model at max reasoning. Up-and-left is better. Kimi K3.1 Pro (red) sits on the efficiency frontier with Gemini 4 Argon; the Claude 5.5 pair are smarter but ~3× the task cost. GPT-6.1 Sol is the budget outlier.
Sources: AA Index + task costs as compiled Oct 2026 (Opus $5.98, Sonnet $5.46, Fable $7.63, Astra $3.26, Argon $1.99, Sol $0.72 per task; K3.1 Pro $1.98 per AA launch article). List prices /1M tokens (in/out): Kimi $2/$10 (cache $0.20) · Opus $4/$20 · Sonnet $2/$10 · Fable $10/$50 · Astra $10/$50 · Argon $2/$10 intro → $4/$20 after 30 Nov · Sol $2/$10. Paradox to note: K3.1 token prices fell 33% vs K3, yet cost/task stayed flat because output grew 48K → 55K tokens/task (+14%). Cheaper tokens ≠ cheaper jobs if the model talks more.
Specs for context (not capability): K3.1 Pro is a 2.8T-total / 104B-active MoE with 1M-token context, text+image input. A Highspeed variant runs 2–3× faster generation (tool time unchanged) at ~2× plan allowance. None of this is evidence of quality — it is evidence of deployability: a near-frontier open-weight MoE changes the economics of self-hosting, fine-tuning, and routing even where it trails by a few points.
Drag the weights to match the work you actually care about. The estimator blends five capability buckets (anchored to the benchmarks above) into a weighted gap vs the best Western model, then converts points to months at a frontier velocity you control.
Presets first, then fine-tune. Weights are normalised automatically.
Blend note: bucket gaps live on different 0–100 scales. To blend them, each bucket is first expressed as a fraction of the best Western score, then mapped onto AA-Index-equivalent points (×58). ProgramBench −33.5/65.0 = −52% relative — that is why strict-completion weight dominates quickly. This is a decision aid, not a measurement.
—
Weighted gap: — Index-equivalent pts · Kimi-equivalent score — vs 58
Negative = pulls Kimi behind. Watch strict completion: a little weight goes a long way.
Try it: Coding-agent shop → ~0 months. Ship-to-spec → 8+ months. Same model, same benchmarks — different buying question. If your answer swings that much, do not quote a single “China gap” number without naming the workload.
Read this before quoting any number from this page. Each item below can move the apparent gap by more than the 3-point headline.
Moonshot ran K3.1 Pro mostly in the Claude Code harness; rival scores are the maximum each vendor reported in blogs/leaderboards, under different harnesses (Codex for Astra on OfficeQA, etc.), effort settings (“max” vs “max with fallback”), and subsets. Harness choice alone can move agent scores several points. §2–§3 inherit all of this. §1 and §4 (Artificial Analysis) are the only same-harness comparisons here — weight them more for rankings, and treat vendor ±2-pt differences as ties.
Kimi Code Bench 3.0, Hallu Bench Hard, and xdev-bench-hard are Moonshot’s own. Hallu Bench is explicitly a stress test (“do not represent overall error rates in daily use”). BenchCAD with-tools scores in the 90s are near ceiling — small gaps there may reflect test saturation, not spatial-reasoning parity. Conversely ProgramBench binary scoring (≥95% or zero) amplifies differences that partial-credit scoring would shrink. Always ask: raw or binary? subset or full?
This page compares frontiers. The median deployed model, inference capacity, chip access, and product distribution are different questions where the US lead is larger. Likewise “China” here means Moonshot on most rows — but DeepSeek leads DeepSWE, and Qwen/GLM/MiMo each lead somewhere. A country-level claim needs the portfolio, not one lab’s best week. K3.1 Pro weights were also not yet released at writing (due by 4 Nov 2026): the “top open-weight model” claim is conditional on shipment, license, and third-party reproduction.
Terminal-Bench 13% → 67% in one post-training round proves these evals respond fast to targeted RL. That cuts both ways: Western labs can (and do) do the same, and benchmark gains may overstate transfer to your private tasks. Contamination, judge-model choice (e.g. Gemini judges in some Moonshot footnotes), and context-compaction settings further blur cross-vendor rows. For a procurement decision, run 20–50 of your tasks in your harness — this page tells you which 20 to pick (long-horizon vs strict-finish vs closed-book), not the final answer.