Kimi K3.1 Pro Early Tests: Top Open-Weights, 70% Cheaper, but Weak on Robotics

After Moonshot released Kimi K3.1 Pro on October 24, the evaluation community put it through a battery of tests the same day. LMArena data shows the model scored 1692 on the Code Arena: WebDev leaderboard, ranking 8th — a 9-place jump over Kimi K3's 17th (1655 points), tying Claude Opus 5 (Max) and lifting Moonshot to 4th on the vendor leaderboard. Pricing is $2/$10 per million input/output tokens, with a median cost of $1.23 per task, roughly 30% of Opus 5's.

Confirmed

• Leaderboard results: 8th place on Code Arena: WebDev (1692 points), tying Claude Opus 5 (Max) at about 70% lower cost; Kimi K3 previously ranked 17th (1655 points).
• Pricing and efficiency: official pricing is $2/$10 per million tokens; cedricchee tested the voxel garden task using the Kimi Code harness — high tier: 5 minutes, 1.06 million tokens, $0.553; max tier: 11 minutes, 1.67 million tokens, $0.95 — overall 2-3x faster than K3, with the "overthinking" issue reportedly fixed.
• 3D/visual generation: Peter Gostev's series of tests covered 3D scenes, HTML animations, websites and UI building; blaind0 generated a "pelican riding a bicycle" 3D scene from a single prompt, beating Opus 5 but losing narrowly to Opus 5.5; ChrisGPT also ran this classic test in about 45 minutes and declared it saturated, and made a trailer in Claude Code, saying visual quality was slightly better than Opus 5.5 though general white-collar tasks still lagged.
• Reasoning benchmarks: a nonobench test relayed by Reddit user fairydreaming shows kimi-k3.1-pro-max became the first known Chinese model to solve a 20x20 nonogram (scoring 10/10 on 5x5).
• Cost comparison cases: on the same "whale breaching the surface" task, Kimi spent $0.75 versus $3.01 for Opus 5.5 and roughly $2+ for GPT-6 Astra.

Not Yet Confirmed

• Open-source timeline: Nathan Lambert noted that if the weights are released as scheduled on November 4, K3.1 Pro's benchmark results would make it the strongest open-weights model to date; he also praised Kimi's explicit cyber risk policy — not letting open weights push the absolute frontier of cyberattack capabilities. Some netizens speculate the training data comes from a specific vendor, but this remains community conjecture.
• Weaknesses under dispute: robocurve data claims K3.1 Pro scored only 9% on real-world robotics tasks, trailing Opus 5.5, Gemini 4 Argon and GPT-6 Astra; Peter Wildeford joked it "forgot to distill robotics capabilities"; researcher Afinetheorem believes the team is strong, but like recent heavily RL-based post-trained models, its general and emergent abilities fall short. Feedback from Chinese communities relayed by teortaxesTex says 3D modeling shows promise, but Kimi Code still draws criticism.

Why It Matters

K3.1 Pro matches Opus 5's WebDev results at roughly 30% of the cost; if it's open-sourced on schedule on November 4, it will significantly reshape the price-performance landscape between open-weights models and frontier closed models. Meanwhile, its shortcomings on embodied tasks like robotics once again expose the ongoing debate over the generality of heavy RL post-training.

2026-10-24 ~ 2026-10-24 · 22 related posts

Full story(7 episodes)→

• Episode 1: Kimi K3.1 leaks pile up: Pro version rumored with open weights by Nov 4(2026-10-22, 18 posts)
• Episode 2: Moonshot teases Kimi K3.1 after strong Arena showing(2026-10-23, 4 posts)
• Episode 3: KOL Predicts Kimi K3.1 Benchmarks Before Testing, Suspects Benchmark Gaming(2026-10-23, 2 posts)
• Episode 4: Kimi K3.1 Pro Rolls Out: 3T-Parameter Flagship with 1M Context(2026-10-24, 12 posts)
• Episode 5: Kimi K3.1 Pro Early Tests: Top Open-Weights, 70% Cheaper, but Weak on Robotics(2026-10-24, 22 posts)
• Episode 6: Moonshot Releases Kimi K3.1 Pro, a 3T-Parameter Open-Weight Model Rivaling GPT-6 and Claude(2026-10-24, 50 posts)
• Episode 7: Moonshot's Kimi K3.1 Pro Hits #4 on Artificial Analysis Index, Closing the Open-Weights Gap(2026-10-24, 19 posts)

• Community tests Kimi K3.1: 3D modeling shows promise, Kimi Code a lost cause — teortaxesTex · 2026-10-24
• [source] Nathan Lambert: Kimi K3.1 Pro looks like the strongest open-weight model, with deliberate cyber pacing — natolambert · 2026-10-24
• Kimi K3.1 Pro hailed as strongest open-weight model amid data-vendor speculation — dejavucoder · 2026-10-24
• Moonshot Kimi now ranks #4 lab on Code Arena WebDev leaderboard — arena · 2026-10-24
• Kimi K3.1 Pro hits #4 lab on Code Arena WebDev, 74% cheaper than Claude Opus 5 — arena · 2026-10-24
• Peter Gostev demos 3D generation with Kimi K3.1 Pro — arena · 2026-10-24
• Kimi K3.1 Pro hits #8 on Code Arena WebDev, matching Claude Opus-5 at 70% lower cost — arena · 2026-10-24
• First Kimi K3.1 Pro benchmarks compare it against Opus 5.5, Sonnet 5.5 and Gemini 4 — arena · 2026-10-24
• Kimi K3.1 Pro takes on the pelican-on-a-bicycle 3D test: beats Opus 5, trails 5.5 — blaind0 · 2026-10-24
• Kimi K3.1 Pro reportedly first Chinese model to solve 20x20 nonograms — fairydreaming · 2026-10-24
• Kimi k3.1 pro aces the pelican-bike test in 45 minutes; author calls benchmark saturated — ChrisGPT · 2026-10-24
• Kimi K3.1 called open-weight Opus 5.5 rival at $0.75 vs $3.01 per task — SumitGup · 2026-10-24
• Kimi K3.1 Pro hands-on: beats Opus 5.5 on visuals, still lags on general-purpose work — ChrisGPT · 2026-10-24
• Hands-On With Kimi K3.1Pro: Open 2.8T-Param Flagship Halves Price, Matches GPT on 3D — 数字生命卡兹克 · 2026-10-24
• Kimi K3.1 Pro tested: 2-3x faster at $0.55 per task, open models close on frontier 3D — cedric_chee · 2026-10-24
• [source] Kimi K3.1 Pro cost breakdown: high mode $0.55/5min, max mode $0.95/11min — cedric_chee · 2026-10-24
• Kimi K3.1 Pro hands-on: 2-3x faster, 5x cheaper than Astra, fixes overthinking — cedric_chee · 2026-10-24
• Kimi K3.1 Pro scores just 9% on real-robot tasks, trailing Opus 5.5 and Gemini 4 Argon — peterwildeford · 2026-10-24
• Kimi K3.1 Pro beats rivals on benchmarks but struggles on robotics and hard tasks — Afinetheorem · 2026-10-24

3 near-duplicate retellings: testingcatalog · cedric_chee · cedric_chee