Claude Mythos 5 for stock trading

Mythos 5 is documentation, not a recommendation. Anthropic released it alongside Claude Fable 5, and both were suspended on 2026-06-12 under a US export-control directive. Existing keys cannot call it. This page exists so operators whose agents ran on Mythos during the window it was live can record accurately what was under the hood, and so anyone reading a trade log from that period knows what they are looking at.

The specs, as published: one million tokens of context, multimodal input, and pricing of $10 per million input tokens and $50 per million output. Mythos and Fable share the same underlying weights. The difference is deployment posture. Fable shipped to the broad API audience with tighter guardrails and automatic fallback to Opus 4.8 on restricted prompt categories. Mythos targeted a narrower, vetted audience with looser restrictions on the same model. For a trading workload neither distinction mattered much, since almost nothing a market agent asks for trips those gates. The pricing did matter.

At $50 per million output tokens, Mythos was the wrong model for a trade loop and a defensible one for offline work. The pattern that made economic sense was research and strategy generation: point the million-token window at a long history of your own fills, your written reasoning, and whatever research you had accumulated, let it produce or revise the strategy logic, then compile that logic into a cheaper model that actually runs the decisions minute to minute. Operators who wired frontier pricing directly into a per-minute scan discovered the bill quickly.

The suspension is open-ended and no return date has been announced. If your agent was on Mythos when it went dark, Opus 4.8 is the closest available substitute: the most capable Claude reachable through the standard API after the suspension, at half the price on both input and output, with a 200K window instead of 1M. If your prompts fit comfortably under 200K, that swap costs you very little. If you were genuinely using the million-token context, you will need to summarize upstream or move to a different provider's long-context model.