Seven days, seven launches, and one clear theme: the AI race is moving on every front at once. Mistral unveiled a 1-trillion-parameter flagship, Meta finally launched its Muse-branded models, Google unveiled a universal agent for the workplace, OpenAI gave ChatGPT interactive answers, Anthropic slashed small-model prices by roughly 75%, an open-weight 501B model landed on Hugging Face, and Google's long-decode Gemini 4 Argon went generally available. Here is what happened, with every claim verified against primary sources.
Key Takeaways
Mistral Large 4 (Oct 6): roughly 1 trillion parameters, trained on 4,000 Nvidia Grace Blackwell GPUs in Europe, with downloadable weights promised on Oct 27. (VentureBeat)
Meta launched two Muse models (Oct 12): standard Wonder at $4/$20 per million tokens, and Wonder Full, available only to Muse Code subscribers. Neither ships downloadable weights. (VentureBeat)
Google unveiled a universal Gemini agent for work (Oct 8) where "coworker agents" get their own Workspace accounts and orchestrate both Gemini and Claude models. (Unite.AI)
OpenAI rolled out GPT-6 to all ChatGPT users with an "Intelligent UI" of interactive answers, while Anthropic's Claude Haiku 5.5 cut token prices to a tenth of Haiku 4.5. (MacRumors, Reuters)
Reflection AI's Beam, a 501B-parameter open-weight model under Apache 2.0, shipped this week and scores 81.6% on SWE-bench Verified. (Hugging Face)
1. Mistral Goes Big: Large 4 "Le Chonk"
The French lab Mistral announced Mistral Large 4 on October 6, a natively multimodal model with roughly 1 trillion total parameters, about 52 billion of them active per token, according to VentureBeat's coverage. It is codenamed "Le Chonk," and it was trained on 4,000 Nvidia Grace Blackwell GPUs inside Mistral's own European data centers.
Two details matter most. First, the preview is available now through Mistral's API, and downloadable weights are promised on October 27 under a custom Mistral license, which would make it one of the largest open-weight models ever released. Second, Mistral's preliminary benchmarks put it at 62% on DeepSWE v1.1 and 15% on Harvey's Legal Agent Benchmark, the latter a legal-agent test where specialized evaluation keeps reshuffling the leaderboard.
The company context is equally loud: Mistral closed a 3 billion euro Series D at a 21 billion euro post-money valuation in September, with a science team of about 300 people. Europe finally has a frontier-scale lab with its own silicon-scale training fleet.
2. Meta's Muse Wonder Enters the Chat
On October 12, Meta's Superintelligence Labs, led by former Scale AI CEO Alexandr Wang, announced two models: Muse Wonder and Muse Wonder Full, VentureBeat reports. This is a sharp break from Meta's Llama era: both models are proprietary, with no downloadable weights.
The lineup is split by access, not just capability. Standard Wonder is available through a commercial API at $4 per million input tokens and $20 per million output tokens (cached input costs $0.20), with a 1-million-token context window, under the model id `muse-wonder-1.0`. Wonder Full is restricted to Muse Code subscribers on the Power plan ($50/month) and Maximum plan ($200/month).
On Artificial Analysis's Intelligence Index, Wonder Full scores 56, tying Claude Sonnet 5.5 and sitting just behind Claude Opus 5.5 at 58. Standard Wonder scores 52. The flagship features are called "Active Memory" and "Active Focus," and Meta claims benchmark leads on DeepSWE 1.1 (79.0) and SWE-Marathon (57.5).
3. Google's Gemini Agent: One Agent for All Your Work
Google Cloud used its Gemini at Work 2026 event on October 8 to unveil the Gemini agent, billed as a single universal agent for work: knowledge work, Q&A, content creation, and coding, from one prompt box and one API, according to Unite.AI's report.
The detail that made people sit up: "coworker agents" get their own Workspace accounts, including an `@agents.company.com` email address, a calendar, Drive access, and directory presence, acting under their own identity. Under the hood, the agent orchestrates across Google's Gemini models and Anthropic's Claude models, with built-in cost controls and per-project spend caps. It runs persistently in the cloud and works inline across Gmail, Docs, Drive, Sheets, Chat, and Calendar.
The agent launched in private preview, with general availability expected in late October or early November. This is the workplace agent war going direct: one universal agent against OpenAI's always-on approach.
4. OpenAI's GPT-6 Gets an "Intelligent UI"
OpenAI started rolling out GPT-6 to ChatGPT users on October 7, alongside an "Intelligent UI" that lets the assistant answer with interactive elements: graphics, tappable buttons, forms, charts, diagrams, maps, bill splitters, calculators, and games, built from a library of native streamable components, reports Juli Clover at MacRumors.
The rollout reached Plus, Pro, Business, and Enterprise users first on October 7, then expanded to Free and Go tiers on October 8. Paid tiers get GPT-6 Sol while Free and Go get GPT-6 Luna, and the new model is limited to the Chat tab; Codex and Work keep their current models. One internal metric OpenAI shared: GPT-6 Instant starts web-search answers 44% sooner than GPT-5.6 Instant, measuring time-to-first-answer only.
5. Anthropic Slashes Prices With Claude Haiku 5.5
Anthropic launched Claude Haiku 5.5 on October 7, and the headline is price: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens ($0.50/$2.50 above that threshold), Reuters reports. That makes it roughly 75% cheaper on average than Haiku 4.5, though Anthropic notes a new tokenizer uses more tokens per task.
This is the first Haiku with an adjustable effort setting, a low-to-max dial for thinking. The specs: 1-million-token context, 128K maximum output. Anthropic's reported benchmarks are punchy: 72.4% on OSWorld 2.1 offline (versus 15.7% for Haiku 4.5), 39.2% on Terminal-Bench 4.0, and 1620 on GDPval-AA v2.1, beating the same-priced GPT-6 Luna on every shared benchmark. It is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and all Claude plans.
6. Reflection AI's Beam: 501B Open Weights
The open-weight camp had its own big week. Reflection AI, a Brooklyn startup backed by Nvidia, Sequoia, and Lightspeed, announced Beam on October 5, and the weights landed this week on Hugging Face under the Apache 2.0 license, per the official Hugging Face blog.
The specs: 501.3 billion total parameters with 23 billion active per token, 52 layers, 256 experts per MoE layer with 8 routed per token, text in and out, and a 1-million-token context window. Team-reported benchmarks: 81.6% on SWE-bench Verified, 97.7% on AIME 2026, and 80.1% on Terminal-Bench v2.1. The NVFP4 checkpoint is about 280GB, which fits on 4x B200 GPUs, and Beam ships with day-0 support in transformers, vLLM, and SGLang plus its own Rust harness.
If you want quick video breakdowns of trending open-source AI projects as they ship, browse the repo signals for 30-second overviews.
7. Gemini 4 Argon Goes Generally Available
Google's Gemini 4 Argon went generally available on October 13, after its September 30 announcement and an initial vetted rollout to cyber defenders, according to kie.ai's analysis, which links to the official announcement and Artificial Analysis profile.
Argon's defining spec is a unified 1,048,576-token input-output budget with long decode up to 1 million output tokens, far beyond the typical 64K-256K output limits elsewhere. Introductory pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026, then rises to $4/$20 from December 1, with cached input at 95% off. It is available through the paid Gemini API, the Gemini app and Search for Google AI Pro and Ultra subscribers, and Antigravity for Ultra.
On Artificial Analysis's Intelligence Index it scores 53, tying GPT-6 Astra. Google reports 77.9% on DeepSWE v1.1 and a 15% hallucination rate on the AA-Omniscience evaluation, which is notably low. It is proprietary, with no weights release.
How the Week's Models Stack Up
Aspect | Model | Price (in/out, per 1M) | Context | Openness | AA Index |
|---|---|---|---|---|---|
Biggest parameter count | Mistral Large 4 | Not disclosed | Not disclosed | Weights promised Oct 27 | Not yet indexed |
Meta flagship | Muse Wonder Full | Muse Code sub only | 1M | Closed | 56 |
Meta standard | Muse Wonder | $4 / $20 | 1M | Closed | 52 |
Google long-decode | Gemini 4 Argon | $2 / $10 (intro) | 1M in/out unified | Closed | 53 |
Budget champion | Claude Haiku 5.5 | $0.10 / $0.50 | 1M | Closed | Not yet indexed |
Open-weight option | Reflection AI Beam | Self-hosted | 1M | Apache 2.0 | Not yet indexed |
Prices are per the cited sources at publication time. The Artificial Analysis Intelligence Index thread across the week: Claude Opus 5.5 (58) leads, Muse Wonder Full (56) ties Claude Sonnet 5.5 (56), then Gemini 4 Argon (53) ties GPT-6 Astra (53), ahead of standard Muse Wonder (52).
Frequently Asked Questions
What is Mistral Large 4 and when do the weights come out?
Mistral Large 4 is Mistral's 1-trillion-parameter flagship model, natively multimodal, announced October 6, 2026. Downloadable weights are promised on October 27 under a custom Mistral license, which would make it one of the largest open-weight models ever released.
How much does Meta's Muse Wonder cost?
Standard Wonder costs $4 per million input tokens and $20 per million output tokens through a commercial API, with cached input at $0.20. Wonder Full is not sold by the token: it is restricted to Muse Code subscribers on the $50/month Power plan and $200/month Maximum plan.
Is Reflection AI's Beam really open source?
Beam is released under the Apache 2.0 license with weights on Hugging Face, which makes it genuinely open for use, modification, and commercial deployment. The weights landed in mid-October 2026, with day-0 support in transformers, vLLM, and SGLang.
What changed in ChatGPT this week?
ChatGPT started getting GPT-6 on October 7, first for Plus, Pro, Business, and Enterprise, then for Free and Go on October 8. The headline feature is "Intelligent UI": answers that include interactive graphics, buttons, forms, charts, and small working tools. Paid tiers get GPT-6 Sol; Free and Go get GPT-6 Luna.
Which model scored highest this week?
Among this week's launches, Muse Wonder Full ties Claude Sonnet 5.5 at 56 on Artificial Analysis's Intelligence Index, the highest of the new releases. Claude Opus 5.5 still leads overall at 58, but it launched in September.
Conclusion
This week compressed a month of AI news into seven days. Europe got its trillion-parameter contender in Mistral Large 4, Meta finally put its Muse brand on frontier models (closed, this time), Google made the workplace agent its next battleground, and the price floor for serious AI fell again with Haiku 5.5. The open-weight ecosystem answered too, with Beam's 501B Apache-licensed weights proving that frontier-scale models do not have to live behind closed APIs.
Read this article in Bahasa Indonesia: Berita AI: 12-19 Oktober 2026.