The LLM landscape fractured this week into two distinct camps: high-performance flagships and ultra-cheap agentic workhorses. Anthropic launched Claude Fable 5.2 and Mythos 5.2 on November 9, claiming superior coding and reasoning capabilities over its predecessor Opus 5.5 while cutting latency by up to 30%. Simultaneously, Google and Meta released budget-focused models designed to handle high-volume tasks at a fraction of the cost, effectively bifurcating the market for developers choosing between raw intelligence and operational efficiency.

Anthropic’s Flagship Leap

Claude Fable 5.2 is generally available via API under the ID claude-fable-5-2, priced at $8 per million input tokens and $40 per million output tokens. The model introduces an optional "fast mode" that generates tokens up to 2.5x faster, albeit at double the cost ($16/$80). Reports indicate it beats Claude Opus 5.5 across most evaluated capabilities, with specific gains in vulnerability discovery for the Mythos 5.2 variant, which also showed significantly fewer safeguard blocks in early testing compared to Fable 5.1.

Google and Meta Go Cheap

Google DeepMind is aggressively positioning itself in the mid-tier with Gemini 4 Carbon, now open to all developers after being restricted to cybersecurity defenders. With introductory pricing at $2/$10 per million tokens rising to $4/$20 on November 30, it’s a clear land grab. Meanwhile, Gemini 4 Hydrogen targets the "daily driver" segment, offering 1M-token context and near-frontier coding scores at $1.50/$7.50 per million tokens. Meta’s Muse Ember is the most radical cost-cut, priced at just $0.10/$0.50 per million tokens, scoring nearly as well as larger models on agentic tasks while running on a 1M-token context window.

The Rise of Decision APIs

Beyond raw generation, a structural shift is occurring in how we handle structured output. Both Hugging Face and Google launched dedicated "decision model" interfaces this week. These APIs bypass the fragility of prompting general LLMs for JSON, instead returning structured probabilities for yes/no, choice, or scoring tasks directly. Google’s Gemini Decision 1 API even supports multimodal input for classification, costing just $0.10 per million input tokens with free output. This suggests the industry is realizing that not every production call needs to be a conversation; many are just complex routing decisions.

Key Takeaways

  • Claude Fable 5.2 offers a drop-in upgrade with 30% faster outputs and lower latency than 5.1.
  • Gemini 4 Hydrogen and Meta’s Muse Ember are explicitly trading slight capability drops for order-of-magnitude cost savings.
  • Decision APIs from Hugging Face and Google aim to replace fragile prompt-based JSON extraction with native structured output endpoints.
  • Gemini 4 Carbon’s price jumps from $2/$10 to $4/$20 on November 30, closing a three-week promotional window.

The Bottom Line

The era of paying frontier prices for routine tasks is ending. If your team isn't already evaluating Muse Ember or Gemini 4 Hydrogen for high-volume agentic workloads, you are likely overpaying by an order of magnitude for negligible capability gains.