Models

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on Oct. 12. Wonder is on the Meta Model API at $4/$20 per million tokens; Wonder Full is for Muse Code subscribers only. Meta says its tests put them alongside Claude Fable 5.1, Opus 5.5 and GPT-6 Astra; Artificial Analysis scores Wonder Full at 56, behind Claude Opus 5.5 at 58. Meta's full safety report is promised but not out.

OpenAI said on Oct. 8 that Ultrafast, its fastest service tier, is rolling out for GPT-6.1 Sol in the API, Codex and ChatGPT Work, at up to 8x the speed of Sol Standard. API pricing is $12 per million input tokens and $60 per million output tokens, which OpenAI calls 1.2x the cost of Astra. The API tier is open to all API users with separate rate limits; in Codex and ChatGPT Work it is limited to Pro 500, eligible usage-based Enterprise and credit-based Edu plans. The speed figure is OpenAI's own.

Anthropic's Claude Haiku 5.5 costs $0.10/$0.50 per million input/output tokens for prompts up to 100,000 tokens, 90% below Haiku 4.5 and level with GPT-6 Luna's API list price; above 100k it is $0.50/$2.50. Artificial Analysis scores Haiku 5.5 at 43 and Luna at 38 on its Intelligence Index, both at max effort. Anthropic's own benchmark lead over Luna is vendor-reported and swings with effort: Terminal-Bench is about 39% at max, about 20% at the default medium setting. Anthropic also cut Sonnet 5.5 cache reads to $0.10.

Mistral launched a paid Studio preview of Mistral Large 4 (ML4 / "le Chonk") on Oct. 6: a natively multimodal MoE with about 1.05T total / 49B active parameters, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its European data centres. Mistral says it will release the weights by the end of the month; TechCrunch and The Next Web put that closer to Oct. 27. Artificial Analysis's first independent Intelligence Index score for the preview is 38 (#64 of 225), behind top open-weight models MiMo-V2.6-Pro (46), GLM-5.3 (45) and Kimi K3 (44). Most other benches are Mistral's own on a checkpoint it says will change.

Reflection published Beam on Oct. 5: a 501-billion-parameter sparse MoE (23B active) open-weight text model it says approaches GLM-5.2 on coding and reasoning with 3–4× less estimated inference compute. Weights, a technical report and a model card are promised later this month under Apache 2.0. Benchmarks are Reflection’s own; Kimi K3 and GLM-5.3 still lead on most shared tasks. Beam is not on Artificial Analysis yet.

Aleph Alpha’s October 3 release, Kolibri, is an English-German mixture-of-experts model with 78.1 billion total parameters and 3.46 billion active, context up to about 1 million tokens (the model card recommends serving at or below 262,144), and full weights on Hugging Face under Apache 2.0. The company says it was trained in Germany and Finland and that 21.3% of pre-training tokens are German. Scores such as AIME 2025 English 96.9 are Aleph Alpha’s own harness at high reasoning effort, not an Arena or Artificial Analysis placement.

Microsoft AI launched MAI-Transcribe-2-Streaming — real-time speech-to-text in 60 languages with continuous language detection — claiming #1 on Artificial Analysis for both final and first-partial transcript accuracy, with first partials in just over 100ms of audio and introductory pricing of $0.54 per hour of audio through year-end. Same day: MAI-Voice-2.1 (23 languages / 26 locales, $22/1M chars) and MAI-Voice-2.1-Flash (~150ms E2E for 45s audio, $15/1M chars). AA ranking is as claimed by Microsoft; treat vendor latency/price figures as company-stated.

Synopsys and OpenAI signed a multi-year preferred-partner agreement to jointly develop GPT-Synopsys, a specialized model trained to operate Synopsys EDA tools, with shared R&D, joint GTM, and an undisclosed revenue framework. Early technology engagements only — no GA date, pricing, benchmarks, or named customers. Customer design data is not used for training.

Google DeepMind’s Koray Kavukcuoglu named Gemini 4 Argon for long-horizon software engineering, legal/finance knowledge work, and cyber defense. Trusted Fairwind defenders and U.S. voluntary pre-release get it without cyber guardrails first; paid API and AI Ultra wait on safeguard iteration. Lab-claimed benches lead DeepSWE and AutomationBench; CWE-bench v1 ties peers at 68%. No independent Arena Elo yet.

OpenAI’s DevDay model is not the shelved Astra successor. GPT-6.1 Sol keeps the $2 / $10 sticker and claims near-GPT-6 Astra results on coding, computer use, and professional work at about one-fifth of Astra’s standard token prices. Those benches are the lab’s. Arena has not ranked it.

Claude Sonnet 5.5 keeps Sonnet 5’s $2/$10 sticker and, at max effort, scores 56 on Artificial Analysis’s Intelligence Index, two points behind Opus 5.5. Anthropic says it costs up to 30% less per task. AA measured about $7.60 a task at max, roughly 50% more than Sonnet 5, because output ran to about 193,000 tokens. Sticker price is not cost per task.

OpenAI halved Sol and Luna API prices versus GPT-5.6 promo rates and put them in Work and Codex the same day. Lab cost-per-task gaps look large; Artificial Analysis still has Opus 5.5 clearly ahead of Sol on intelligence — the trade this launch asks buyers to accept.

Anthropic priced Claude Opus 5.5 at $4/$20 with cache reads at $0.20 — Fable-class safeties, ~40% cheaper to run than Opus 5. Artificial Analysis puts it ahead of GPT-6 Sol on intelligence; Andon’s Vending-Bench cash result does not.

xAI shipped Grok 4.7 at the same $2/$6 token prices as Grok 4.6, with a fast variant at 2× speed and 2× price. Lab tables lead EEBench and Harvey’s legal-agent bench; Andon’s Vending-Bench put it in the middle — closer to Opus 5.5 than to Sol.

xAI shipped Grok Voice Transcribe 2.0 at unchanged $0.10 batch / $0.20 streaming per hour. Lab short-phrase WER falls from 20.6% to 6.8%; xAI cites first place on Artificial Analysis’s streaming accuracy board among 32 models.

At Dreamforce, Salesforce and NVIDIA launched Koa, Salesforce’s first CRM reasoning model for Agentforce, post-trained from Nemotron 3 Super on a synthetic 27-year CRM corpus. Salesforce keeps the weights. Vendor CRM Benchmark claims 3× fewer errors; pilot now, GA winter 2026 US.

Google shipped Gemini 3.8 Live and Live Extended Thinking for production voice agents. Extended Thinking tops Artificial Analysis’ Speech-to-Speech Quality Index at 82.6; API audio is $0.005/min in and $0.018/min out. SynthID on all generated audio.

OpenAI launched GPT-6 Astra, the first model it says meets Preparedness Critical cybersecurity, plus a $1 billion Daybreak for Frontline Defenders program. Lab benches saturate FrontierMath Tier 4 and ARC-AGI-3; independent Arena/AA listings were still pending at launch.

Meta Superintelligence Labs released Muse Spark 1.3 — about 20% fewer tool calls and 25% fewer tokens than 1.2 on internal coding. Artificial Analysis Intelligence 61 at xhigh versus Fable 5.1 max at 66. Fourth Spark in five months.

Google launched Gemini 3.8 Flash at the same $0.75/$3.75 intro price as 3.7 through year-end, plus 3.8 Flash Cyber for trusted defenders via the new Fairwind program. Public Flash stays the workhorse; Critical-tier tools stay gated.

Anthropic launched Claude Fable 5.1 generally and Mythos 5.1 for trusted-access, cut cache reads 75% to $0.25/MTok, and opened Enterprise Frontier Safeguards so monitoring data stays in the customer’s cloud. Same week OpenAI called Astra Critical cyber.

Google launched Gemini Omni 1.1 Flash for developers: scene extension to 40 seconds from 10 seconds of prior context, first/last-frame interpolation, 360p drafts, 4K upscale, and three-second video references. Production video SKU of Gemini Omni — not a new Flash language model. Customers cited: Adobe Firefly, Figma Weave, Runway.

Google launched Gemini 3.5 Transcribe — streaming and file speech-to-text with 4.0%/2.6% Artificial Analysis WER, 70% faster time-to-final than Chirp 3, 85+ languages, and live APIs already in Gboard Rambler on Pixel 11. Independent AA WER is the check; noisy call-center audio is the open question.

Alibaba’s Qwen team open-sourced Qwen3.8-Flash-Next — a 125B multimodal MoE with 6B active parameters, 262K native context, and vendor scores that preview the Qwen4 architecture — at $0.16/$0.47 per million tokens on the production Flash API. Independent Arena / Artificial Analysis numbers were not confirmed — label vendor scores as lab-claimed.