AI Model & Benchmark Watch — November 6, 2026

A quiet week at the frontier: Kimi K3.1 Pro's weights finally went public, Artificial Analysis nudged its top scores up a point (Gemini 4 Carbon 60, Claude Opus 5.5 59), and the launches that did happen were mostly gated, niche, or still unverified.

Overview

Last week's loose end closed. Moonshot's Kimi K3.1 Pro now has a Hugging Face repository under a custom license, which makes the 46 it scores on Artificial Analysis (AA) the best result by a downloadable model, 14 points ahead of DeepSeek V4.1 Flash 1017's 33. At the top, AA's leaderboard now shows Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59, one point above what we recorded last Friday; the page doesn't say why. New model news was thin: OpenAI's GPT-6 Rosalind is a restricted life-sciences model, Meituan's LongCat-2.5 has an API but no confirmed weights, and Qwen 4 still hasn't shipped.

New & Updated Models (this week)

Commercial / closed

GPT-6 Rosalind — OpenAI (reported October 30, 2026)

Who / License: OpenAI; closed, restricted to approved organizations through OpenAI's trusted-access program.

What's notable: A follow-up to April's GPT-Rosalind life-sciences model, aimed at drug discovery and research workflows. I could not find OpenAI's own announcement for this exact name, so the details come from secondary write-ups. One of them, citing the system card addendum, says it uses the same weights as GPT-6 Astra with no extra training; that claim rests on a single source. Aggregators list $15 / $75 per million tokens, also unconfirmed. Not in the table below because it is not generally available and has no AA score.

Source: TechJack Solutions summary · The Decoder — OpenAI launches GPT-Rosalind · CloudPrice — GPT-6 Rosalind

LongCat-2.5 — Meituan (reported November 2, 2026)

Who / License: Meituan; API and chat service for now. Open weights are unconfirmed: one report says they are planned, another says Meituan hasn't committed, and a BAAI hub post says no 2.5 repository exists on GitHub or Hugging Face (LongCat-2.0, MIT, is still the latest listed).

What's notable: Coverage is secondary and thin (one outlet describes a 1.6T-parameter multimodal model with a 1M context window). I did not verify the specs or any benchmark claims, so none are repeated here. Watch for a repository.

Source: AF.NET — Meituan launches LongCat-2.5 · BAAI hub · nocode.mba — Meituan

Open-weight

Kimi K3.1 Pro — weights released — Moonshot AI

Who / License: Moonshot AI; open weights on Hugging Face under the custom "Kimi K3.1 Pro License" (AA: commercial use allowed with restrictions). The model card gives no release date; Moonshot had promised weights by November 4.

What's notable: 2.8T total parameters, 104B active (Mixture-of-Experts), 1,048,576-token context, text and image input, MXFP4 weights. AA scores it 46 and lists API pricing at $2 / $10 per million tokens. Moonshot's model card compares it with Claude Opus 5.5: Terminal-Bench 4.0 65.7 vs 66.4, GPQA-Diamond 94.2 vs 93.9, HLE-Full 51.1 vs 59.8, OSWorld V2.1 79.6 vs 81.8. Those are Moonshot's numbers, several from internal benchmarks or custom harnesses, and I haven't seen them reproduced. Note the 65.7 on Terminal-Bench is lower than the 67.2% AA's leaderboard showed last week; the two come from different harnesses.

Source: Hugging Face — moonshotai/Kimi-K3.1-Pro · Artificial Analysis — Kimi K3.1 Pro

Smaller or unconfirmed items

• Tencent Hy4: one release tracker lists an "Hy4" announcement on November 4 with weights and system cards due "in the coming days." Every other source I found dates the open Hy4 Preview (770B total, 49B active, Apache 2.0) to August 28, so treat the November entry as unconfirmed. nocode.mba tracker · DataNorth — Hy4 preview
• Qwen 4: still no release. A report says an Alibaba executive told an October 29 audience it arrives "next week"; other coverage treats that as unverified, and Alibaba hasn't said whether any weights will be public. Emergent — Qwen4 lineup · Latenode
• DeepSeek V4 Flash 0731: LLM Gateway lists it with a November 5 date, but other sources describe a July 31 release, so I'm not counting it as new. LLM Gateway timeline · DataNorth

Head-to-Head — Current Frontier

AA Intelligence Index, context, and prices were checked on AA model pages and the AA leaderboard on November 6. For multi-effort models the table uses the highest-effort variant AA lists. Columns marked ‡ are carried from the October 30 edition and were not re-pulled this week: LMArena Elo (aggregator-reported, arena.ai blocks direct fetch), SWE-bench Pro, and GPQA Diamond. The Kimi GPQA figure is Moonshot's own. "—" means no current sourced figure.

Prices marked "per Oct 30 edition" were sourced from AA model pages in that edition and not re-fetched today.

State of play: Carbon leads on AA but is reachable only by vetted cyber-defense partners, so Opus 5.5 remains the best score you can actually call. Among models you can download, Kimi K3.1 Pro (46) is now a real option, not a pending one, and it sits 13 points below Opus 5.5 and 14 above the next open model. Moonshot's own card has it within a point of Opus on GPQA and Terminal-Bench but 8.7 points behind on HLE-Full, which is the sort of gap an aggregate index smooths over. On price per capability, Kimi at $2 / $10 matches Sonnet 5.5's list price with a lower AA score, while DeepSeek V4.1 Flash 1017 at $0.30 / $1.20 is the cheap open option.

Benchmark & Leaderboard Movement

• Every top score moved up about a point. AA's leaderboard now shows Carbon at 60 (we recorded 59), Opus 5.5 at 59 (58) and Sonnet 5.5 at 55 (54). The page carries only "Updated," so I can't say whether this is a re-run, a new variant listing, or a methodology tweak. Compare week-to-week numbers cautiously.
• Ordering is unchanged at the top: Carbon, Opus 5.5, Astra, then Sol and Sonnet 5.5 tied at 55.
• Kimi's open-weight lead is now real. Last edition flagged that the best "open" score belonged to a model with no public weights. That's resolved: 46 versus DeepSeek V4.1 Flash 1017's 33 (AA lists it at 33; last week we had 32) is a 13-point gap between the two best open models.
• Self-reported claims are piling up. Moonshot's model card benchmarks, Google's cyber evals for Carbon, and the Hy4 numbers all come from the labs themselves. None of the figures in this edition's table depend on them except the labelled Kimi GPQA cell.

Analysis

For closed-model work, Opus 5.5 is still the highest-scoring model you can call without an invitation, with Sonnet 5.5 and GPT-6.1 Sol tied two notches below at 55 for less money. If you need weights on your own hardware, Kimi K3.1 Pro is the new top of the pile, but 2.8T parameters means this is a cluster deployment, not a workstation one; DeepSeek V4.1 Flash 1017 is the practical choice and the cheapest to run through an API. For the open-versus-closed gap, the headline is 13 to 14 points on AA between the best open and best generally available closed model, narrower than the 27 we calculated last week against a downloadable model, and Qwen 4 could move it again if it lands. I'd wait on LongCat-2.5 and Hy4 until someone publishes a repository and independent scores.

Sources

• Artificial Analysis — Leaderboard
• Artificial Analysis — Gemini 4 Carbon
• Artificial Analysis — Claude Opus 5.5
• Artificial Analysis — Claude Sonnet 5.5
• Artificial Analysis — GPT-6 Astra
• Artificial Analysis — Kimi K3.1 Pro
• Artificial Analysis — DeepSeek V4.1 Flash 1017
• Artificial Analysis — GLM-5.3
• Artificial Analysis — Mistral Large 4
• Hugging Face — moonshotai/Kimi-K3.1-Pro
• TechJack Solutions — GPT-6 Rosalind summary
• The Decoder — OpenAI launches GPT-Rosalind
• CloudPrice — GPT-6 Rosalind
• AF.NET — Meituan launches LongCat-2.5
• BAAI hub — LongCat-2.5
• nocode.mba — Meituan release tracker
• nocode.mba — November 2026 release tracker
• LLM Gateway — model timeline
• llm-stats — AI news
• DataNorth — Tencent releases Hy4 preview
• DataNorth — DeepSeek V4-Flash-0731
• Emergent — Alibaba teases Qwen4 lineup
• Latenode — Alibaba teases Qwen4 rollout
• My-Library — AI Model & Benchmark Watch, October 30, 2026 edition