The hierarchy of LLM capabilities has shifted. Between October 13 and October 16, 2026, the industry saw a simultaneous release of Google’s closed-source Gemini 4 Argon and two massive open-weight models from Meta and Reflection. This week proved that 'frontier-class' is no longer a proprietary exclusive, as open weights now match closed models in parameter count, context length, and agentic reasoning capabilities.

The Open-Weight Heavyweights

Meta dropped 'Muse Spark Open,' an 816B-parameter mixture-of-experts model that the company claims is the most capable US open model to date. It features 42B active parameters per token and a 1M-token context window, specifically tuned for agentic coding. Just one day later, Reflection, a new US lab, released 'Beam,' a 501B-parameter MoE with 23B active parameters and Apache 2.0 licensing. Both models are optimized for long-horizon engineering tasks, signaling that open weights are now the default choice for serious agentic workflows.

Google’s Pricing Gambit

Google’s Gemini 4 Argon went to general availability on October 13, targeting complex multi-step work in coding, finance, and legal. The introductory pricing is aggressive: $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95%. However, this is a limited-time offer; on November 30, 2026, prices will double to $4/M input and $20/M output. This move puts Gemini 4 Argon at parity with Claude Opus 5.5, forcing developers to lock in usage now to avoid a sudden cost spike.

Specialized Tools and Infrastructure

While the giants fought for parameter dominance, smaller releases focused on efficiency and specialization. Mistral introduced 'Timestral,' a family of time-series foundation models ranging from 5M to 209M parameters that run entirely on CPU, eliminating the need for GPU infrastructure for forecasting tasks. IBM’s Granite Switch 4.1 bundled chat, decision, and speech-to-text adapters into a single endpoint, reusing KV cache to keep memory costs low. Meanwhile, LightOn’s ColGrep-Agent offers a 2B-parameter local code-search model that cuts API token usage by 34% for coding agents.

Key Takeaways

  • Open weights have achieved parity with closed frontier models in both size and capability.
  • Google’s pricing for Gemini 4 Argon will double on November 30, 2026.
  • Specialized, CPU-only models like Mistral’s Timestral are reducing infrastructure barriers for niche tasks.
  • The 'agentic coding' use case is now the primary battleground for both open and closed models.

The Bottom Line

The era of proprietary superiority in raw scale is over; developers must now weigh open-weight control against the ticking clock of Google's introductory pricing.