[punitarani]

Requested by Punit · project thread

Opened by the nightly ModelBeats sync routine. Draft: Punit reviews and merges.

Before: Kimi K3.1 Pro's only Terminal-Bench Science number is Moonshot's own 38.1 (Claude Code harness). Gemini 4 Argon, Claude Haiku 5.5 and Granite 4.2 8B have no Text Arena rating, because Argon was still Preliminary and the other two weren't on the board the corpus last read (Oct 13).

After: Vals AI's independent Terminal-Bench Science run for Kimi K3.1 Pro (31.43) is the headline. The three models carry their ratings from Arena's Oct 27 board. Gemini 4 Argon moves to #1 (3156.6), 0.8 ahead of Kimi K3.1 Pro.

Corrections (4 rows)

Model · benchmark	Before	After	Source
Kimi K3.1 Pro · Terminal-Bench Science 0.1	38.1 self-reported	31.43 independent (±5.59, #8 of 43)	TB-Science board, updated Oct 25 (added Oct 28)
Gemini 4 Argon · Text Arena	none (1525 Preliminary on Oct 13)	1520 ±6, 11,889 votes	Text Arena, Oct 27 (added Oct 29)
Claude Haiku 5.5 · Text Arena	none	1425 ±7 (High), 7,890 votes	same (added Oct 29)
Granite 4.2 8B · Text Arena	none	1290 ±11, 3,251 votes	same (added Oct 29)

• Kimi K3.1 Pro. Vals ran it through Moonshot's API at max reasoning effort with mini-swe-agent. Vals credits no provider fallbacks for this run, so the fallback rule doesn't apply. The Oct 26 and 27 runs read a cached board copy dated Oct 21 and missed it.
• Arena rows. These follow the Oct 13 convention: add rows only for corpus models that had none. Carbon (1528) is Preliminary, as are Gemini 3.8 Flash (High), Gemini 3.7 Flash (High) and Nemotron 3.5 Super VL, so none of them are filed. Existing rows are point-in-time snapshots and stay as they are; most moved by 6 points or less since Oct 13 (for example Muse Wonder 1513 → 1507).

Leaderboard moves (both commits, against main)

	Before	After
Gemini 4 Argon	#2 (3126.1)	#1 (3156.6)
Kimi K3.1 Pro	#1 (3169.9)	#2 (3155.8)
Muse Wonder Full	#3 (3100.5)	#3 (3102.3)
Claude Opus 5.5	#4 (3072.4)	#4 (3065.9)

• Claude Haiku 5.5 drops from #33 to #49 (2593.2 → 2450.2). Its 1425 is far below the frontier models it is otherwise rated against.
• Granite 4.2 8B gets its first scored row and stays unranked.
• Most other ratings move by about 5 points or less.

Counts

597 models (unchanged), 6329 → 6333 results.

Re-pins

• Goldens: derive top 5 and feat(data): add the DeepMind Gemini + Poolside Laguna releases since Kimi K3 #1 slug (Gemini 4 Argon), and movers (Sarvam 105B +1620.8).
• e2e:
  • Rankings, explorer, dashboard rail, quick-compare default and closed-frontier leader now show Gemini 4 Argon (3156.6).
  • Movers is +1620.8, the compare legend is GPT-5.2 2382.9 and DeepSeek-V3.1 (Thinking) 1811.2, and the Opus 4.6 family delta is +241.9.

Checked, nothing to change

• Vals: Terminal-Bench 4.0 (updated Oct 27, 57 rows) matches the corpus apart from Gemini 4 Carbon's new row, which feat(data): ingest Gemini 4 Carbon and Particle-1, Mistral Large 4 open weights (Oct 26–31, 2026) #25 files. The rows that aren't filed are the documented exclusions. Terminal-Bench Science (still Oct 25) matches too.
• ARC Prize: no new results since DeepSeek V4.1 Flash 1017, which the corpus has.
• Pending weights: Mistral Large 4's weights are still not confirmed. Hugging Face wasn't reachable from this run. ActuIA's check on the evening of Oct 28 found the repo still showing an upcoming release with an Oct 31 ETA and no files or license. One blog says the weights are out, but nothing first-party backs it up.

Merge order with #25

#25 (Gemini 4 Carbon and Particle-1) touches the same derived files and goldens. Whichever merges second needs a re-derive and re-pin. The next nightly run does that on whichever PR is still open.

Gates

• generate-dataset, derive and validate-data all pass, and so does bun run ci.
• Playwright: 115 of 115 pass on the current head.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CAthXVhtrPcUij6nbrXMv7