Anthropic skips ARC-AGI-3 for Claude Mythos

The "Mythos" omission confirms that ARC-AGI-3 is the new "vibes check" for AGI, and right now, nobody is passing.

• –While Mythos sets records on SWE-bench (93.9%) and USAMO (97.6%), these are still "pattern-matching" domains compared to ARC's interactive reasoning.
• –The 0.37% state-of-the-art score (Gemini 3.1 Pro) suggests that current LLM architectures struggle with novel rule discovery in game-like environments.
• –Anthropic’s silence likely stems from marketing; even a "good" 1% score looks like a failure to a non-technical audience used to 90%+ benchmark numbers.
• –Project Glasswing's restricted access makes independent verification impossible, allowing Anthropic to curate the model's performance narrative.
• –The extreme token costs ($125/M output) suggest Mythos is optimized for high-value reasoning, making its potential failure on basic ARC tasks even more telling.