News 2026-10-23

AI Model & Benchmark Watch — October 23, 2026

The top of the leaderboard didn't move this week; the middle got crowded.

AI Model & Benchmark Watch — October 23, 2026

The top of the leaderboard didn't move this week; the middle got crowded. Reflection AI shipped its first model, DeepSeek pushed a six-point update that ties a Chinese rival for the #2 open-weight spot, and Anthropic and OpenAI both shipped agent products built on existing models rather than new ones.

Overview

Claude Opus 5.5 is still at 58 on the Artificial Analysis Intelligence Index, Muse Wonder Full and Claude Sonnet 5.5 are still tied one step behind at 56, and nothing challenged either figure this week. The action was one tier down. Reflection AI, a startup that spent three years in stealth before its first release, put out Beam, a 501B-parameter open-weight model that lands at 30 on the Index — a real debut, but one that trails Meta's own open release from a week ago (39) and even the UAE's K2 Horizon (31). DeepSeek's quieter move did more damage to the standings: V4.1 Flash 1017 picked up six points and now sits at 45, tied with GLM-5.3 for the second-best open-weight score, one point behind MiMo-V2.6-Pro. Anthropic and OpenAI both shipped agent products this week — Claude Duet and parallel agents in GPT-6 Pro — rather than new underlying models, which is its own kind of signal about where the competition is headed next.

New & Updated Models (this week)

Commercial / closed

Claude Duet (Research Preview) — Anthropic (October 21, 2026)

Who / License: Anthropic; closed, available through a waitlist on Pro and Max plans in the US.

What's notable: Duet isn't a new model — it runs on Claude Opus 5.5 — but it's a new way of using one. It's a standalone app (iOS, Android, claude.ai) where Claude works proactively: it sends a morning brief across your connected apps, keeps working after you close the tab, and messages you only when something needs a decision. It connects to more than 6,000 tools through the Claude Marketplace and runs on its own virtual machine with a dedicated browser, so it doesn't need your laptop open. Anthropic requires explicit approval before Duet sends a message, makes a payment, signs anything, or deletes content, and blocks it from trading securities or moving money between accounts even with approval. Anthropic calls this "an early version" and says it can still misjudge urgency.

Source: Unite.AI — Anthropic Opens Claude Duet Research Preview to US Pro and Max Users · Engadget — Claude Duet is Anthropic's answer to Meta Muse and OpenAI Dots

GPT-6 Pro — Parallel Agents — OpenAI (October 16, 2026)

Who / License: OpenAI; closed, rolling out to eligible ChatGPT Pro subscribers.

What's notable: Also not a new model — this is a capability added to the existing GPT-6 Pro. A single request can now be split across several agents working at once, each one visible in the conversation as it runs, with results stitched together at the end. OpenAI's own example is market research split three ways (competitors, pricing, regulation) instead of run in sequence. No independent benchmark accompanies the launch since the underlying model hasn't changed.

Source: TokenPost — OpenAI Adds Parallel Agents to GPT-6 Pro and Updates Codex Usage · CellCog — GPT-6 Pro in ChatGPT Gets Parallel Agents

Open-weight

Beam — Reflection AI (October 16, 2026)

Who / License: Reflection AI; Apache 2.0. Weights, tokenizer, and reference implementation are on Hugging Face, with day-0 support in transformers, vLLM, and SGLang.

What's notable: This is Reflection's first shipped model, after the company spent three years building before releasing anything. Beam is a 501B-parameter mixture-of-experts model with 23B active parameters, a 1M-token context window, and a design goal of using fewer reasoning tokens per task than its rivals — Artificial Analysis measured about 26,000 output tokens per Intelligence Index task, roughly half what K2 Horizon uses for a similar score. That efficiency claim holds up better than the overall intelligence claim: Beam scores 30 on the Index, independently verified by Artificial Analysis, which puts it #2 among non-Chinese open-weight models (behind Meta's Muse Spark Open at 39, just ahead of K2 Horizon at 31) but well below the Chinese open leaders at the top of this week's table. It's also notably weak on Terminal-Bench 4.0 (3.0%, behind even Qwen3.8 27B's 5.6%) while doing well on AA-Omniscience (33% correct, ahead of K2 Horizon's 18%) and AutomationBench-AA (48.5%, ahead of K2 Horizon's 37.2%). Priced at $0.40 input / $1.20 output per million tokens.

Source: Reflection AI — Beam is now available · Artificial Analysis — Reflection has released Beam, their first model launch · Artificial Analysis — Beam 1.0 · OfficeChai — Reflection AI Releases Beam, Becomes Second-Best US Open Model

DeepSeek V4.1 Flash 1017 — DeepSeek (October 17, 2026)

Who / License: DeepSeek; MIT license, open weights.

What's notable: A version-numbered update to DeepSeek V4.1 Flash, and a bigger jump than the name suggests: the Intelligence Index score rose from 39 to 45, a six-point gain that ties GLM-5.3 for the second-best open-weight score on this week's board, one point behind MiMo-V2.6-Pro. The gains concentrate in reliability rather than raw reasoning — its non-hallucination rate on AA-Omniscience jumped from 4% to 33%, and its GDPval-AA Elo rose 146 points (1600 to 1746) — which is a different kind of improvement than a new architecture, but a real one. Context window holds at 1.05M tokens; pricing is unchanged at $0.30 input / $1.20 output per million tokens.

Source: Artificial Analysis (X) — DeepSeek V4.1 Flash 1017 scores 45 · OfficeChai — DeepSeek V4.1 Flash 1017 Places Ahead Of GLM 5.3, Kimi K3 · Artificial Analysis — DeepSeek V4.1 Flash 1017

Ling-3.1-flash-1019 — InclusionAI / Ant Group (October 22, 2026)

Who / License: InclusionAI; open weights "coming soon" per the announcement — not yet downloadable as of this writing, though the updated version is already live via API.

What's notable: Another version bump, not a new base model: the Index score moved from 41 to 44, tying Kimi K3, on the strength of agentic-task gains — AA-Briefcase Elo rose from roughly 1400 to 1524, and AutomationBench-AA from 62% to 68%. Worth flagging the same caveat we'd apply to any "open" model whose weights haven't shipped yet: the score is real and independently measured, but you can't run it yourself until InclusionAI actually publishes the files. Context window is 1M tokens, text in and out.

Source: Ant Ling (X) — Introducing Ling-3.1-flash-1019 · OrcaRouter — Ling-3.1-flash-1019 Ties Kimi K3 at 44

MiniMax M3.1 Flash — MiniMax (October 19, 2026)

Who / License: MiniMax; open weights.

What's notable: A 428B mixture-of-experts model aimed at agentic workloads, priced at $0.15 per million input tokens. MiniMax's own numbers include 73.7 on DeepSWE v1.1 and 82.2 on CyberGym, but Artificial Analysis's independent read puts it at 42–43 on the Intelligence Index — close to but behind DeepSeek V4.1 Flash 1017 (45) — and flags a high hallucination tendency on top of that score. Not strong enough to crack this week's head-to-head table below.

Source: MarkTechPost — MiniMax Just Launched M3.1 Flash · Miraflow — MiniMax M3.1 Flash Explained: Benchmarks, Pricing, and the Hallucination Problem

Still waiting: Mistral Large 4's open weights are targeted for "by the end of the month" per Mistral's own post; some outlets report October 27, Hugging Face's placeholder page says October 31 — treat the later date as the one Mistral has actually committed to. Its API preview, running since October 6, scores 38 on the Index. Qwen 4 remains in training, and GLM-5.5 is unannounced.

Head-to-Head — Current Frontier

Standings as of October 23, 2026. AA Index is Artificial Analysis's Intelligence Index v4.3.2, data snapshot October 20. LMArena Elo comes from aggregator reporting on the arena.ai leaderboard; arena.ai itself returned a 403 to direct fetch this week, so treat these as a recent but not freshly re-pulled snapshot, same caveat as last week. "—" means a current, sourced figure could not be confirmed.

Model	Org	Open?	AA Index (v4.3.2)	LMArena Elo	GPQA Diamond	SWE-bench (Verified/Pro)	Terminal-Bench 4.0	Context	$ / Mtok in/out
Claude Opus 5.5	Anthropic	No	58	1507 (high)†	—	89.9% (Pro)*	59.6%	1M	$4 / $20
Muse Wonder Full	Meta	No	56	—	—	—	57%*	1M	— (no public API)
Claude Sonnet 5.5	Anthropic	No	56	—	—	—	63.6%	1M	$2 / $10
Claude Fable 5.1	Anthropic	No	53	1501 (max)†	92.6%	81.2% (Pro)*	57.9%	1M	$10 / $50
GPT-6 Astra	OpenAI	No	53	—	96.1% (max)	—	58.2%	1M	$10 / $50
Gemini 4 Argon	Google DeepMind	No	53	1525 (high)†	—	—	—	1M	$2 / $10
GPT-6.1 Sol	OpenAI	No	52	1484 (max)†	—	—	—	1.05M	$2 / $10
Grok 4.7	xAI	No	46	—	—	—	37.6%	500K	$2 / $6 (<200K)
MiMo-V2.6-Pro	Xiaomi	Yes (MIT)	46	—	—	—	—	1M	$0.43 / $0.87
GLM-5.3	Zhipu / Z.ai	Yes (custom license, MaaS clause)	45	—	—	77.8% (Verified)	41.8%	1M	$1.40 / $4.40
DeepSeek V4.1 Flash 1017 (new)	DeepSeek	Yes (MIT)	45	—	—	—	—	1.05M	$0.30 / $1.20
Kimi K3	Moonshot AI	Yes (custom license)	44	1488 (max)†	93.5%*	93.40% (Verified)	—	1.05M	$3 / $15
Ling-3.1-flash-1019 (new)	InclusionAI	Yes (weights pending)	44	—	—	—	—	1M	— (API only)
Claude Haiku 5.5	Anthropic	No	43	—	—	—	—	1M	$0.10 / $0.50 (≤100K)
Gemini 3.8 Flash	Google DeepMind	No	41	1497 (high)†	95.3%	—	19.7%‡	1M	$0.75 / $3.75
Muse Spark Open	Meta	Yes (Muse Model License)	39	—	—	—	—	1M	$1.25 / $4.25
K2 Horizon 375B-A23B	MBZUAI / IFM	Yes (open weights)	31	—	—	—	—	—	—
Beam (new)	Reflection AI	Yes (Apache 2.0)	30	—	—	—	3.0%	1M	$0.40 / $1.20

* Reported by the vendor or at launch, not independently reproduced on the current benchmark version. † Aggregator-reported arena.ai figure, not independently re-pulled this week. ‡ Scored after the September 7 Terminal-Bench 4.0 hardening pass.

Reading the table: Nothing changed at the very top — Opus 5.5 (58), Sonnet 5.5's Terminal-Bench lead (63.6%), and GPT-6 Astra's GPQA lead (96.1%) are all identical to last week. The movement is in the open-weight pack between 39 and 46, where three updates landed in five days: DeepSeek's 1017 revision jumped six points to tie GLM-5.3, Ling-3.1-flash-1019 jumped three points to tie Kimi K3 (though its weights aren't downloadable yet), and Reflection AI entered the board entirely with Beam. None of those three unseated MiMo-V2.6-Pro, still the single best open-weight score at 46, and none of them came close to Muse Spark Open's 39 among non-Chinese open models, let alone the Chinese leaders above it.

Benchmark & Leaderboard Movement

DeepSeek's quiet update outranked Reflection's loud debut. V4.1 Flash 1017's six-point jump (39 → 45) moved it further up the board than Beam managed on a first release — a reminder that an incumbent's point update can outweigh a new entrant's launch.
Reflection AI shipped, after three years of stealth. Beam is a real open-weight debut (Apache 2.0, 501B/23B active), but it lands at 30 on the Index — third among non-Chinese open models, behind Muse Spark Open (39) and K2 Horizon (31), and far below the Chinese open leaders.
"Open" is getting a looser definition. Ling-3.1-flash-1019 posted an independently measured score before its weights exist anywhere to download. That's a score you can cite, not a model you can run — worth separating the two when a lab claims an "open" release.
Two labs shipped agent products, not models. Claude Duet and parallel agents in GPT-6 Pro both run on existing frontier models. Neither moves a leaderboard number, but both are bets that the next competitive axis is what you build on top of a model rather than the model itself.
No movement at the very top, for the second week running. Opus 5.5 (58), Sonnet 5.5's Terminal-Bench lead (63.6%), and GPT-6 Astra's GPQA lead (96.1%) are unchanged.

Analysis

For agentic coding, Sonnet 5.5 still holds the one public, independently run number that matters most (Terminal-Bench 4.0 at 63.6%); Muse Wonder Full's vendor-reported coding claims from last week are still unverified on that same benchmark. For reasoning, GPT-6 Astra still holds GPQA. If you need to self-host, MiMo-V2.6-Pro (46) is still the single best open-weight score, but the gap below it just got more competitive: GLM-5.3 and DeepSeek V4.1 Flash 1017 are tied one point back, and Kimi K3 and Ling-3.1-flash-1019 are tied one point behind that — four real options in a two-point band, before you even get to Reflection's new entrant. The open-versus-closed gap at the top held exactly where it was (46 vs. 58); what changed is how many open-weight teams are now clustered just below that ceiling, shipping updates within days of each other.

Sources

Open raw folder ↗ News/AI-News/AI_Model_Benchmark_Watch/October_2026/model-benchmark-watch_2026-10-23

Metadata

Overview

Title	AI Model & Benchmark Watch — October 23, 2026
Category	News
Date	2026-10-23
Slug	model-benchmark-watch-2026-10-23
Description	The top of the leaderboard didn't move this week; the middle got crowded.
Tags	model-benchmark-watch
Source	News/AI-News/AI_Model_Benchmark_Watch/October_2026/model-benchmark-watch_2026-10-23