Satire. Every score on this site is made up. The tweets are real.Read the disclaimer →
Trust Me Bro
Get VerifiedContribute
ModelsBenchmarksIncidentsIn the WildMethodologyChart Crime Studio
TMB-2026-0017··Alleged benchmaxxing
🤝🤝🤝Disputed

The most benchmaxxed model ever released, independently confirmed to score well on benchmarks

Subject: Kimi K3.1 Pro

Claim

Moonshot's launch thread: 65.7% on Terminal-Bench 4.0, "ahead of GPT-6 Astra"; "SOTA on OfficeQA Pro v2"; and, "on internal benchmarks", "~4.2× K3 on simulated business operation". Fan summary: it "beats every listed US frontier model on 6 of 14 benchmarks in Moonshot's evaluations".

Finding

Sixty-six minutes after release it was declared "the most benchmaxxed model ever released". Hands-on reports followed: worse than Sonnet 5.5 for one user, beaten by DeepSeek V4.1 Flash at 49 times the price in one test, eleventh on a human-preference arena, 21% on BiGym 2. The complication: Artificial Analysis independently measured 67% on Terminal-Bench 4.0, above Moonshot's own 65.7% and the highest of any model, and an index of 55, fourth of 215. The suspicious part is the shape. Terminal-Bench 4.0 rose from 13% to 67% in one version; Humanity's Last Exam rose five points. That is either a breakthrough in agents or a breakthrough in Terminal-Bench, and we are not qualified to say which. Judging by the replies, neither is anyone else.

Timeline

  1. 23 Oct 2026
    Self-reported: "ahead of GPT-6 Astra" on Terminal-Bench 4.0.@Kimi_Moonshot · ♥ 2.1K ↗
  2. 23 Oct 2026
    "On internal benchmarks … ~4.2× K3 on simulated business operation."@Kimi_Moonshot · ♥ 770 ↗
  3. 23 Oct 2026
    Independent: 67% on Terminal-Bench 4.0, up from 13% for Kimi K3, the highest of any model.@ArtificialAnlys · ♥ 260 ↗
  4. 23 Oct 2026
    "The most benchmaxxed model ever released."@The_Alex · ♥ 914 ↗
  5. 23 Oct 2026
    "Even Sonnet 5.5 gives me better results."@pankajkumar_dev · ♥ 112 ↗
  6. 24 Oct 2026
    Eleventh on Agent Arena, "below even Opus 5 and Fable 5".@thegenioo · ♥ 147 ↗
  7. 25 Oct 2026
    "Couldn't even beat DeepSeek V4.1 Flash on xhigh. At 49x the price."@EnvolDev · ♥ 166 ↗
  8. 25 Oct 2026
    21% on BiGym 2.@Nag1ovo · ♥ 48 ↗
  9. 26 Oct 2026
    "Paid 29x more for the worse one."@semichenkko · ♥ 17 ↗

Primary sources

The posts this entry is based on, embedded as published.

Kimi_Moonshot
Kimi.ai
@Kimi_Moonshot
Link to tweet
On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%).
On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%). https://t.co/qh5wU1qAX4
♥ 2.1KOpen on X ↗
Kimi_Moonshot
Kimi.ai
@Kimi_Moonshot
Link to tweet
K3.1 Pro is SOTA on OfficeQA Pro v2 (64.4). On internal benchmarks, it improves over K3 by 7.2 points on financial modeling and 11.6 points on financial research, and scores ~4.2× K3 on simulated business operation. Example outputs: an NVDA DCF/LBO model with fully linked
K3.1 Pro is SOTA on OfficeQA Pro v2 (64.4). On internal benchmarks, it improves over K3 by 7.2 points on financial modeling and 11.6 points on financial research, and scores ~4.2× K3 on simulated business operation. Example outputs: an NVDA DCF/LBO model with fully linked https://t.co/M8mZWjq8yw
♥ 770Open on X ↗
ArtificialAnlys
Artificial Analysis
@ArtificialAnlys
Link to tweet
Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro
Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro https://t.co/OAszElzDGo
♥ 260Open on X ↗
The_Alex
Alex
@The_Alex
Link to tweet
Kimi k3.1 is the most benchmaxxed model ever released
♥ 914Open on X ↗
pankajkumar_dev
Pankaj Kumar
@pankajkumar_dev
Link to tweet
I feel like Kimi K3.1 is benchmaxxed because I am not getting outputs anywhere near the level shown in the benchmarks. Even Sonnet 5.5 gives me better results, and Opus 5.5 is on another level.
♥ 112Open on X ↗
thegenioo
Hamza
@thegenioo
Link to tweet
Agent Arena is one of the benchmarks I actually am starting to trust bcz it is built on real people using these models for such tasks Kimi K3.1 Pro is #11 there, below even Opus 5 and Fable 5 from last gen, which kind of says it is benchmaxxed in terms of claiming that it is
Agent Arena is one of the benchmarks I actually am starting to trust bcz it is built on real people using these models for such tasks Kimi K3.1 Pro is #11 there, below even Opus 5 and Fable 5 from last gen, which kind of says it is benchmaxxed in terms of claiming that it is
♥ 147Open on X ↗
EnvolDev
Tony
@EnvolDev
Link to tweet
Kimi K3.1 Pro vs. DeepSeek V4.1 Flash vs. MiniMax M3.1 Flash ⚔️ Everyone says Kimi K3.1 Pro is beating Opus 5.5 on some benchmarks. In my test, on MAX, it couldn't even beat DeepSeek V4.1 Flash on xhigh. 👀 At 49x the price. 🤯 Kimi: $24.12 DeepSeek: $0.49 MiniMax: $0.23 Am
♥ 166Open on X ↗
Nag1ovo
Jesse Zhang
@Nag1ovo
Link to tweet
Kimi 3.1 Pro scores 21% on BiGym 2, behind GPT-6 Astra, Claude Opus 5.5 and GPT-6.1 Sol. bigym2.github.io/leaderboard/
♥ 48Open on X ↗
semichenkko
Semichenko
@semichenkko
Link to tweet
i asked minimax and kimi for the same meteor strike and paid 29x more for the worse one 29x price gap top - minimax m3.1 flash, $0.21. debris bounces, embers float up, the treeline burns one tree at a time bottom - kimi k3.1 pro, $6.10. chunky cubes, a smoke column with no
♥ 17Open on X ↗