News 2026-10-30

AI Model & Benchmark Watch — October 30, 2026

Artificial Analysis tore up its own ruler this week — the Intelligence Index relaunched as v5.0 with harder, mostly-private tests — and every score in this edition's table moved as a result, hours…

AI Model & Benchmark Watch — October 30, 2026

Artificial Analysis tore up its own ruler this week — the Intelligence Index relaunched as v5.0 with harder, mostly-private tests — and every score in this edition's table moved as a result, hours before Google's Gemini 4 Carbon claimed the new #1 spot anyway.

Overview

Artificial Analysis rebuilt its Intelligence Index from the ground up on October 27, swapping in two new benchmarks, upgrading three more, retiring three others, and pushing 60% of the test set onto data nobody outside AA has seen. Every absolute score in this week's table is a product of that new ruler, not a continuation of last week's numbers — Claude Opus 5.5 led the relaunch at 58, only to be passed a day later by Google's Gemini 4 Carbon at 59, a model most people still can't use. Moonshot's Kimi K3.1 Pro posted the best score of any "open" model at 46, except its weights won't exist until November 4 — the same pattern we flagged with Ling-3.1 last week, now repeating with a bigger number attached. Mistral, meanwhile, did the unglamorous thing and actually shipped: Large 4's weights landed on Hugging Face on October 29, three weeks after its API preview opened.

New & Updated Models (this week)

Commercial / closed

Gemini 4 Carbon — Google DeepMind (October 28, 2026)

Who / License: Google DeepMind; closed, staged rollout starting with vetted cyber-defense partners through the Fairwind program (650+ organizations). No public release date yet.

What's notable: Google calls Carbon its strongest frontier model, built specifically for cybersecurity work and long-horizon coding — it arrives just two weeks after Gemini 4 Argon and scores nine points higher on the relaunched Intelligence Index (59 vs. Argon's 50). Google's own testing puts it at 88.2% on SEC-bench Pro (vulnerability-finding), 54.8% on ExploitGym, and 71.0% on CWE-bench (patching) — all vendor-reported, not independently reproduced. On the one neutral benchmark it has been measured on, Terminal-Bench 4.0, it scores 65.2%, second to Kimi K3.1 Pro's 67.2%.

Source: RDWorld Online — Google's Gemini 4 Carbon Takes Benchmark Lead · Google — Gemini 4 Carbon: our next frontier model · The Decoder — Google DeepMind releases Gemini 4 Carbon

GPT-6.1 Terra — OpenAI (October 26, 2026)

Who / License: OpenAI; closed. Live in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users; available via API as gpt-6.1-terra.

What's notable: A fast, mid-tier model built for ordinary computer work — reading documents, filling out forms, matching product photos to catalog entries — not frontier reasoning. OpenAI says its rate of factual errors at low reasoning effort fell from 53.6% to 16.7% versus GPT-5.6 Terra, and it scores 50.3% on Terminal-Bench 4.0 at maximum effort against 21.5% published for its predecessor. Context window is 1.05M tokens (922K input, 128K output), and the API costs $1 input / $5 output per million tokens — half of GPT-5.6 Terra's price.

Source: Decrypt — OpenAI Launches GPT-6.1 Terra, Its Half-Good AI Model at Half the Price · Qz — OpenAI is launching cheaper mid-tier GPT-6.1 Terra

Genie 4 — Google DeepMind (October 27, 2026)

Who / License: Google DeepMind; closed research preview for trusted testers, with a wider rollout to Google AI Ultra subscribers and Gemini API partners planned "in the coming weeks."

What's notable: Not a chat model and not on the table below — Genie 4 is a world model that generates real-time interactive scenes with native audio (music, ambient sound, dialogue) and physics simulation, holding a consistent scene for up to two minutes, double Genie 3's limit. Gemini acts as a "director," planning steps and checking frames while Genie renders them. Google cautions that visual consistency still degrades over longer sessions.

Source: TechCrunch — Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering · Unite.AI — Google DeepMind Rolls Out Genie 4 World Model in Research Preview

Open-weight

Kimi K3.1 Pro (+ Highspeed) — Moonshot AI (October 23–24, 2026)

Who / License: Moonshot AI; weights not yet public — full release promised by November 4. 1M-token context, native image and video input, priced at $2 input / $10 output per million tokens, a third cheaper than K3.

What's notable: A post-trained upgrade of Kimi K3 with substantially more post-training compute, not a new base model. It scores 46 on the relaunched Intelligence Index v5.0 — the highest of any open or pending-open model, 14 points clear of the best model you can actually download today (DeepSeek V4.1 Flash 1017 and GLM-5.3, tied at 32). It also tops the standalone Terminal-Bench 4.0 leaderboard outright at 67.2%, ahead of Gemini 4 Carbon (65.2%) and Claude Sonnet 5.5 (63.6%). Moonshot's own launch comparison puts Claude Opus 5.5 at 66.4% on the same benchmark — that figure is Moonshot's, not independently verified here.

Source: The Decoder — Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5 · OrcaRouter — Kimi K3.1 Pro: $2/$10 Launch and a Tougher Benchmark Sheet · Artificial Analysis — Kimi K3.1 Pro

Mistral Large 4 — weights released — Mistral AI (October 29, 2026)

Who / License: Mistral AI; Mistral Ecosystem Open-Weight License, FP8 and NVFP4 checkpoints on Hugging Face. 1.05T total parameters, 52B active (Mixture-of-Experts), natively multimodal, 1M-token context.

What's notable: The weights follow an API preview that's been live since October 6 — three extra weeks of training before the public release. Scores 26 on the relaunched Intelligence Index v5.0; its 38 under the now-retired v4.3.2 index isn't a fair before/after, since the whole ruler changed underneath it this week, not just Mistral's model. API pricing is $0.70 input / $2.10 output per million tokens.

Source: TechCrunch — Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity · SiliconANGLE — Mistral releases open weights for Mistral Large 4 · Artificial Analysis — Mistral Large 4

Liquid AI LFM3 — Liquid AI (October 27, 2026)

Who / License: Liquid AI; open-weight. Six ~2B-class models — text, vision, audio, omni, and two decision variants — all branched from a shared 1.7B base.

What's notable: Built for on-device deployment, not frontier reasoning, so it doesn't appear in the table below. The text model (LFM3-1.7B) runs in roughly 1GB at 4-bit precision and averages 49.37 across 29 benchmarks — ahead of Alibaba's Qwen3.5-9B (47.24), a model more than five times its size. Worth tracking as its own competitive axis: small-model efficiency, separate from the frontier race.

Source: MarkTechPost — Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models · AIWeekly — Liquid AI opens six-model LFM3 family at 256K context

Still waiting: Qwen 4 remains unreleased; an Alibaba Cloud executive reportedly told Apsara Conference attendees on October 29 it "arrives next week," but nothing is official. GLM's next release and DeepSeek V4.2 have no credible release window.

Head-to-Head — Current Frontier

Standings as of October 29–30, 2026, under Artificial Analysis Intelligence Index v5.0 — a full methodology relaunch on October 27, so these scores are not directly comparable to any prior edition of this table. LMArena Elo comes from aggregator reporting on the arena.ai leaderboard, snapshot dated October 20; arena.ai itself returned a 403 to direct fetch again this week, same caveat as the last two editions. "—" means a current, sourced figure could not be confirmed.

Model	Org	Open?	AA Index (v5.0)	LMArena Elo	GPQA Diamond	SWE-bench Pro	Terminal-Bench 4.0	Context	$ / Mtok in/out
Gemini 4 Carbon (new)	Google DeepMind	No	59	— (gated, not yet on arena.ai)	—	—	65.2%	1M	— (cyber-defender preview only)
Claude Opus 5.5	Anthropic	No	58	1516 (high)†	—	89.9%	—	1M	$4 / $20
GPT-6 Astra	OpenAI	No	57	—	96.1% (max) / 96.3% (xhigh)	—	—	1M	$10 / $50
GPT-6.1 Sol	OpenAI	No	55	—	—	—	—	1M	$2 / $10
Kimi K3.1 Pro (new, weights pending)	Moonshot AI	Pending (Nov 4)	46	—	—	—	67.2%	1M	$2 / $10
Claude Sonnet 5.5	Anthropic	No	54	—	—	81.3%	63.6%	1M	$2 / $10
Muse Wonder Full	Meta	No	50	—	—	—	—	1M	— (no public API)
Gemini 4 Argon	Google DeepMind	No	50	1532 (high)†	—	—	—	1M	$4 / $20
Claude Fable 5.1	Anthropic	No	49	1512 (max)†	—	81.2%	—	1M	$10 / $50
Claude Haiku 5.5	Anthropic	No	36	—	—	—	—	1M	$0.10 / $0.50
Grok 4.7	xAI	No	34	—	—	—	—	500K	$2 / $6
DeepSeek V4.1 Flash 1017	DeepSeek	Yes (MIT)	32	—	—	—	—	1M	$0.30 / $1.20
GLM-5.3	Zhipu / Z.ai	Yes (custom license)	32	—	—	—	—	1M	$1.40 / $4.40
Gemini 3.8 Flash	Google DeepMind	No	31	1497 (high)†	95.3%	—	—	1M	$0.75 / $3.75
MiMo-V2.6-Pro	Xiaomi	Yes (MIT)	30	—	—	—	—	1M	$0.43 / $0.87
Mistral Large 4 (new, weights released)	Mistral AI	Yes (Mistral Ecosystem Open-Weight)	26	—	—	—	—	1M	$0.70 / $2.10
Muse Spark Open	Meta	Yes (Muse Model License)	25	—	—	—	—	1M	$1.25 / $4.25
Beam 1.0	Reflection AI	Yes (Apache 2.0)	18	—	—	—	—	1M	$0.40 / $1.20
K2 Horizon 375B-A23B	MBZUAI / IFM	Yes (open weights)	18	—	—	—	—	524K	~$0 (median across free-tier providers)

† Aggregator-reported arena.ai figure, not independently re-pulled this week.

Reading the table: Gemini 4 Carbon leads at 59, a point above Claude Opus 5.5 (58) — but Carbon is gated to cyber-defense partners, so Opus 5.5 is still the best score anyone outside that program can actually call. Kimi K3.1 Pro's 46 would be the clear best open-weight score, except nobody can download it until November 4; among models you can run today, DeepSeek V4.1 Flash 1017 and GLM-5.3 are tied for the open-weight lead at 32, with Mistral's freshly-weighted Large 4 close behind at 26. The gap between the big labs' open releases (DeepSeek, GLM, MiMo, all 30-32) and the smaller open entrants (Beam, K2 Horizon, both 18) is wider than it looked last week — whether that's the new benchmarks exposing a real capability gap or just reflecting where each team spent its tuning budget isn't something this table can settle.

Benchmark & Leaderboard Movement

Artificial Analysis relaunched its Intelligence Index as v5.0 on October 27. It added FrontierCode 2.0 and Terminal-Bench-Science, upgraded HLE-Diamond, AA-Omniscience, and CritPt v2, retired GDPval-AA v2.1, SciCode, and AA-LCR v1.1, and shifted 60% of the index onto private or held-out test sets. Every score in this week's table is a product of that new mix — treat it as a new baseline, not a continuation of last week's numbers.
The open-vs-closed gap widened on paper, but the real story is which "open" model they used to measure it. AA's own announcement put the gap at 12 points (58 to Kimi K3.1 Pro's 46) at the moment v5 launched. That 46 belongs to a model with no public weights until November 4 — the same move we flagged with Ling-3.1-flash-1019 last week. Measured against a model anyone can actually download today, the gap is 27 points (59 to DeepSeek V4.1 Flash 1017's 32), not 12.
Gemini 4 Carbon took #1 within a day of the index it topped. Claude Opus 5.5 led the v5.0 launch at 58; Carbon's own benchmark run landed hours to a day later at 59 — a reminder that "current #1" on a index this fresh can have a shelf life measured in hours.
Terminal-Bench 4.0 got a new leader, and it isn't Carbon or Opus. Kimi K3.1 Pro's 67.2% is the highest score on that leaderboard this week, ahead of Gemini 4 Carbon (65.2%) and Claude Sonnet 5.5 (63.6%) — the last of those unchanged from last week, since this standalone benchmark wasn't touched by the Index overhaul.
GPQA Diamond and SWE-bench Pro leadership held. GPT-6 Astra is still on top of GPQA Diamond (96.1% max / 96.3% xhigh) and Claude Opus 5.5 is still on top of SWE-bench Pro (89.9%), both identical to last week.
Mistral shipped weights, not just an API. Large 4's Hugging Face release on October 29 is the actual event; its Intelligence Index score moving from 38 (old index) to 26 (new index) is not evidence the model got worse — it's the same methodology shift affecting everyone else in this table.

Analysis

If you need a model you can run today, nothing beats Claude Opus 5.5 (58) among closed options or DeepSeek V4.1 Flash 1017 / GLM-5.3 (32, tied) among open ones — Gemini 4 Carbon and Kimi K3.1 Pro both score higher, but one is gated to a cybersecurity preview and the other has no public weights yet. For coding specifically, Kimi K3.1 Pro's Terminal-Bench 4.0 score (67.2%) is the one to watch once it ships, but Claude Sonnet 5.5 (63.6%) is still the best number anyone can use right now. For reasoning, GPT-6 Astra's GPQA Diamond lead is unchanged. For cheap and fast, Claude Haiku 5.5 ($0.10/$0.50) remains the value pick. The open-vs-closed story this week isn't really about a widening gap — it's that two weeks running, the best "open" score has gone to a model nobody could actually open.

Sources

Open raw folder ↗ News/AI-News/AI_Model_Benchmark_Watch/October_2026/model-benchmark-watch_2026-10-30

Metadata

Overview

Title	AI Model & Benchmark Watch — October 30, 2026
Category	News
Date	2026-10-30
Slug	model-benchmark-watch-2026-10-30
Description	Artificial Analysis tore up its own ruler this week — the Intelligence Index relaunched as v5.0 with harder, mostly-private tests — and every score in this edition's table moved as a result, hours…
Tags	model-benchmark-watch
Source	News/AI-News/AI_Model_Benchmark_Watch/October_2026/model-benchmark-watch_2026-10-30