AI Model & Benchmark Watch — October 9, 2026

The frontier didn't move this week, but the cheap end of the market did: Claude Haiku 5.5 now outscores every budget open-weight model on the Artificial Analysis Index at a lower per-token price, and Mistral and Reflection both announced big open-weight models whose weights you can't download yet.

Overview

Nobody shipped a new flagship. The top of the Artificial Analysis Index is exactly where it was last Friday, with Claude Opus 5.5 at 58 and Sonnet 5.5 at 56. The action was one tier down. Anthropic's Claude Haiku 5.5 (October 7) scores 43 on the Index at $0.10/$0.50 per million tokens. That beats Gemini 3.8 Flash (41) and DeepSeek V4.1 Flash (39), and it costs less per token than either. Mistral Large 4 (October 6) and Reflection AI's Beam (October 5) are the two biggest open-weight stories, but both are announcements for now: Mistral's weights are due at the end of October, Beam's "later this month," and neither can be downloaded today. OpenAI made GPT-6 the default in ChatGPT on October 7, but that's a rollout of the Sol and Luna models it shipped in September, not a new model.

New & Updated Models (this week)

Commercial / closed

Claude Haiku 5.5 — Anthropic (October 7, 2026)

Who / License: Anthropic; closed, available on the Claude Platform, AWS, Google Cloud, and Microsoft Azure as claude-haiku-5-5.

What's notable: Haiku 5.5 is the small model Anthropic previewed alongside Opus 5.5 on September 22, and it was worth the two-week wait. The price is the headline: $0.10 input and $0.50 output per million tokens for prompts up to 100K tokens, rising to $0.50/$2.50 above that. Haiku 4.5 cost $1/$5, so per token this is 90% cheaper on shorter prompts. Anthropic's own estimate is that it costs about 75% less to run on average. It's also the first Haiku with adjustable effort, from Low up through Max. Context is 1M tokens.

Artificial Analysis independently scores it 43 on Index v4.3.2, at $0.21 per Index task and 240.5 output tokens per second. For comparison, DeepSeek V4.1 Flash costs $0.27 per task and Gemini 3.8 Flash costs $1.24. One caveat from the same page: Haiku used 440M output tokens to run the Index, so it's chatty, and the low per-token price is doing a lot of the work.

Anthropic's launch table pits it against Haiku 4.5, OpenAI's GPT-6 Luna, and its own Sonnet 5.5. Anthropic claims 72.4% on OSWorld 2.1 (Luna: 48.9%), 1620 Elo on GDPval-AA v2.1 (Luna: 1437), 39.2% on Terminal-Bench 4.0 (Luna: 16.4%), and 45.9% on Humanity's Last Exam without tools. These are Anthropic's numbers. The Terminal-Bench figure in particular comes from Anthropic's own harness, which last week put Sonnet 5.5 at 70.6% against Artificial Analysis's 63.6%, so don't read it against the public board.

Source: Anthropic — Introducing Claude Haiku 5.5 · Anthropic — Claude Haiku 5.5 System Card · Artificial Analysis — Claude Haiku 5.5 · Let's Data Science — Anthropic releases Claude Haiku 5.5

Mistral Large 4 (Preview) — Mistral AI (October 6, 2026)

Who / License: Mistral AI; API-only preview for now. Artificial Analysis currently lists it as proprietary. Open weights are promised for the end of October.

What's notable: Mistral's first trillion-parameter model. It's a mixture-of-experts with 1.05T total parameters and 49–52B active (the announcement and the model card disagree slightly), plus a 1.6B vision encoder. It takes text and image input and outputs text. Mistral lists a 1M-token context; Artificial Analysis's listing for the preview API shows 524K. List price is $1.36/$4.18 per million tokens, with a preview rate of $0.68/$2.09 according to llm-stats.

Artificial Analysis scores the preview at 38 on the Index, which puts it 20 points behind Opus 5.5 and 8 behind the best open-weight model it tracks. Mistral's own claims are about security work: 93% on Cybench and 82% on a vulnerability reproduction-and-patch test. Le Monde Informatique notes that some of the published results come from tests Mistral ran or chose itself. One newsletter tally puts it at 49.8% on a combined Coding Agent Index and 59.9% on AutomationBench. Mistral says it trained the model entirely in its own European data centers on Grace Blackwell GPUs, which matters to European buyers who care about sovereignty more than leaderboard rank.

Source: Artificial Analysis — Mistral Large 4 Preview · Le Monde Informatique — Mistral dévoile Large 4 · llm-stats — Mistral Large 4 · AI Weekly (Substack) — Mistral Large 4 and the Decision Model Wave

GPT-6 becomes ChatGPT's default — OpenAI (October 7, 2026)

Who / License: OpenAI; closed.

What's notable: This is a product change, not a new model. GPT-6 Sol now serves Plus, Pro, Business, and Enterprise users, and GPT-6 Luna rolls out to Free and Go users starting the following day. Both models launched September 22. The new part is "Intelligent UI," which lets responses include charts, forms, buttons, and small interactive tools, and lets ChatGPT start answering while it's still thinking. No new benchmarks or prices were published. On the leaderboard side, Artificial Analysis added a "GPT-6 Sol (Daybreak Blue)" entry on October 8 and says it leads its Cyber Index. That's the cyber-defender variant OpenAI offers through its Daybreak program, not a general release.

Source: 9to5Mac — OpenAI brings GPT-6 to ChatGPT and debuts Intelligent UI · OpenAI — GPT-6 and Intelligent UI for everyone · Artificial Analysis — Changelog

Open-weight

Beam — Reflection AI (October 5, 2026)

Who / License: Reflection AI; weights promised under Apache 2.0 "later in October." As of October 6 they weren't on Hugging Face, Ollama, or OpenRouter.

What's notable: Reflection's first model, pitched as a Western answer to the Chinese open-weight frontier. It's a text-only sparse MoE with 501B total parameters and 23B active, pretrained for 1M-token context (the API beta caps input at 262K). Reflection claims 80.9% on SWE-bench Verified and 80.1% on Terminal-Bench v2.1, and says Beam matches GLM-5.2 on reasoning while using "3-4x less inference compute." TechCrunch notes that none of this has been independently verified. DataCamp points out that the compute estimate leaves out prompt processing, serving overhead, and long-context attention costs. Reflection's own comparison table also shows Beam behind GLM-5.3, Kimi K3, and DeepSeek V4.1 on Terminal-Bench v2.1 and DeepSWE. Its business plan is less about a public API and more about selling custom "AI factories" to enterprises and governments.

Source: TechCrunch — Reflection AI launches Beam · DataCamp — Reflection AI Beam

Smaller open releases this week: Google put out EmbeddingGemma 2 (October 6, Apache 2.0), a 740M-parameter multimodal embedding model with 8,192-token context. Perplexity released the pplx-embed-v2-late embedding models (0.6B and 9B, MIT) on October 7, and put Apache 2.0 weights out for Perplexity Decider v1 27B, joining Cloudflare's Clef from last week in the small "decision model" trend. None of these is a general chat model, so none goes in the table.

Source: BenchLM — AI Model Releases in October 2026 · AI Weekly (Substack) — Mistral Large 4 and the Decision Model Wave

Still waiting: Qwen 4 is still in training. GLM-5.5 is still unannounced. Gemini 4 Argon is still limited to Google's Fairwind cyber-defense partners, and there's no date for paid API access.

Head-to-Head — Current Frontier

Standings as of October 9, 2026. AA Index is Artificial Analysis Intelligence Index v4.3.2, unchanged since September 19, so the scores compare directly with last week's. LMArena Elo comes from arena.ai's text leaderboard as updated October 8. Terminal-Bench 4.0 shows Artificial Analysis's independently run public-board scores. Vendor "supplemental" numbers are discussed in the model sections above and kept out of the table. "—" means I couldn't confirm a current, sourced figure.

* Reported by the vendor or at launch, and not independently reproduced on the current benchmark version. † Scored after the September 7 Terminal-Bench 4.0 hardening pass.

Reading the table: The top ten didn't change. Opus 5.5 still leads the Index at 58, Sonnet 5.5 still leads Terminal-Bench 4.0 at 63.6%, and GPT-6 Astra still holds GPQA Diamond at 96.1%, with Artificial Analysis's GPQA page showing 96.3% for its xhigh setting. On Arena, Gemini 4 Argon stays #1 at 1525, while Opus 5.5 moved up to 1507 from 1504 as votes came in. The change is at the bottom of the table. Haiku 5.5 is now the cheapest model here per token and outscores Gemini 3.8 Flash and DeepSeek V4.1 Flash on the Index. Until this week, the budget tier was where open-weight models had the clearest edge. MiMo-V2.6-Pro is still the best open model overall at 46, and still cheaper per Index task ($0.13 against Haiku's $0.21), so open models haven't lost on price. But the margin is a lot thinner than it was a week ago.

Benchmark & Leaderboard Movement

• A closed model now leads the budget tier. Claude Haiku 5.5 scores 43 on the AA Index at $0.10/$0.50, ahead of Gemini 3.8 Flash (41, $0.75/$3.75) and DeepSeek V4.1 Flash (39, $0.30/$1.20). Only MiMo-V2.6-Pro is cheaper per Index task.
• Mistral Large 4 debuts at 38 on the Index. That's a respectable number for a European model, but it sits below several open models that already shipped weights, including GLM-5.3 (45) and DeepSeek V4.1 Flash (39). Its real test comes when the weights arrive.
• Opus 5.5 gains on Arena (1504 to 1507, now 6,272 votes) and GPT-6.1 Sol shows up on the board at 1484, well behind the Anthropic and Google leaders. Gemini 4 Argon is unchanged at #1 (1525).
• The flagships held still. No changes in the AA top ten and no new #1 on the AA Index, Arena, Terminal-Bench 4.0, or GPQA Diamond.
• Watch the K2 Horizon numbers. One aggregator reported MBZUAI's fully open K2 Horizon 375B-A23B at 47 on the AA Index, which would have made it the top open model. Artificial Analysis's own page shows 31. MiMo-V2.6-Pro (46) is still the open-weight leader.
• Two open-weight announcements, zero downloadable weights. Both Beam and Mistral Large 4 promise weights by the end of October. For the second week running, no new general-purpose open-weight model you can actually download has shipped.

Analysis

For high-volume work like classification, summarization, support bots, or subagents in a coding loop, Haiku 5.5 is now the first thing to try. It scores higher than the budget models it replaces at a tenth of Haiku 4.5's per-token price. Watch the long-prompt rate ($0.50/$2.50 above 100K) and how verbose it gets. For agentic coding, Sonnet 5.5 still leads, and nothing launched this week challenged it. Reasoning still belongs to GPT-6 Astra on GPQA. If you need weights you can host yourself, MiMo-V2.6-Pro and GLM-5.3 remain the picks. Beam and Mistral Large 4 are worth testing once their weights actually land, but right now that's a promise and a set of vendor benchmarks. On open versus closed, the Index gap hasn't moved (46 vs. 58). What changed is that closed labs are now competing on price, which was open-weight's strongest argument.

Sources

• Anthropic — Introducing Claude Haiku 5.5
• Anthropic — Claude Haiku 5.5 System Card
• Let's Data Science — Anthropic releases Claude Haiku 5.5 with lower prices and effort controls
• 9to5Mac — OpenAI brings GPT-6 to ChatGPT and debuts Intelligent UI
• OpenAI — GPT-6 and Intelligent UI for everyone
• Le Monde Informatique — Mistral dévoile Large 4, un modèle à plus de 1 000 milliards de paramètres
• llm-stats — Mistral Large 4
• AI Weekly (Substack) — Mistral Large 4 and the Decision Model Wave
• TechCrunch — Reflection AI launches Beam
• DataCamp — Reflection AI Beam
• BenchLM — AI Model Releases in October 2026
• LLM Gateway — New AI Models timeline
• Let's Data Science — Google releases Gemini 4 Argon to cyber defenders
• AI Weekly — MBZUAI's IFM ships K2 Horizon
• Artificial Analysis — LLM Leaderboard
• Artificial Analysis — Changelog
• Artificial Analysis — Claude Haiku 5.5
• Artificial Analysis — Mistral Large 4 Preview
• Artificial Analysis — K2 Horizon 375B A23B
• Artificial Analysis — Ling 3.1 Flash
• Artificial Analysis — MiMo-V2.6-Pro
• Artificial Analysis — DeepSeek V4.1 Flash
• Artificial Analysis — Gemini 3.8 Flash
• Artificial Analysis — Claude Opus 5.5
• Artificial Analysis — Claude Sonnet 5.5
• Artificial Analysis — Claude Fable 5.1
• Artificial Analysis — GPT-6 Astra
• Artificial Analysis — GPT-6.1 Sol
• Artificial Analysis — Gemini 4 Argon
• Artificial Analysis — Muse Spark 1.3
• Artificial Analysis — GLM-5.3
• Artificial Analysis — Kimi K3
• Artificial Analysis — Terminal-Bench 4.0 Benchmark Leaderboard
• Artificial Analysis — GPQA Diamond Benchmark Leaderboard
• Arena.ai — Text leaderboard (updated October 8, 2026)
• CodingFleet — Terminal-Bench 4.0 Leaderboard
• My-Library — AI Model & Benchmark Watch, October 2, 2026 edition