This Week in AI: Gemini 4 Argon Goes Wide, Meta Ships Muse Wonder, and the Enterprise Gets Agentic

The second week of October 2026 will be remembered as the week the million-token context window went from a differentiator to table stakes — and the week enterprise agents stopped being pilots and started being plumbing. Here's what matters for people building AI software.

Gemini 4 Argon hits general availability

Google's Gemini 4 Argon is now globally available through the Gemini API, AI Studio, and the Gemini platform after a two-week cyber-defender exclusive. It brings a 1-million-token context window and leads benchmarks including DeepSWE v1.1 and LVBench. Introductory pricing of $2 per million input tokens and $10 per million output runs through November 30 before doubling to $4/$20 — so if Argon is on your eval list, run your harness now while it's half price. It's already wired into GitHub Copilot and Databricks.

Meta enters the closed-weights race with Muse Wonder

Meta released two closed-weight frontier models, Muse Wonder and Muse Wonder Full, both with 1-million-token context windows and a new Active Memory/Focus capability. Muse Wonder is API-accessible at $4/$20 per million tokens (with aggressive $0.20 cached-input pricing), while Muse Wonder Full is gated behind Meta's Muse Code subscription plans at $50–$200/month. Meta charging subscription-gated access to its best weights is a notable strategic turn — watch how the open-weights community reacts.

Claude Haiku 5.5 reprices subagent work

Anthropic's Claude Haiku 5.5 landed on October 7 at $0.10 per million input tokens — a price point clearly aimed at the high-volume verification and subagent layers of agentic architectures. If your workflow runs a cheap model as a checker behind a frontier model (see this week's hallucination deep dive), your verification budget just shrank.

Agents become enterprise plumbing

Three announcements mark the same trend from different angles. Workday integrated its Sana agent with Anthropic's Claude via the Model Context Protocol (MCP) for governed HR and finance tasks — MCP is quietly becoming the standard rail for enterprise agent connections. Microsoft introduced Work IQ, grounding Copilot agents in live Dynamics 365 and Power Platform data. And OpenAI's new 'Dots' product gives Pro and Business users always-on GPT-6 Astra agents capable of persistent, multi-day task execution. The pattern: agents that live inside systems of record, with governance, are where the enterprise money is going.

Quick hits

• JetBrains released Mellum2.1, an open-weights 12B mixture-of-experts coding model (2.5B active parameters), on Hugging Face.
• Character.AI filed for Chapter 11 bankruptcy; Disney has offered a $20M non-exclusive license for its technology.
• Liquid AI shipped two new decision models: d1-3B and the experimental d1-omni-600M.

The takeaway for builders: context is cheap and abundant, frontier choice is widening weekly, and the integration layer — MCP, governed data access, persistent agents — is where the real engineering problems (and advantages) now live. Run your eval harness before you swap models. See you next week.