Claude Fable 5 — Anthropic's First Mythos-Class Frontier Model, Explained

Transcript

Show Host: On June ninth, twenty-twenty-six, Anthropic shipped the most powerful model it has ever made generally available — Claude Fable five, the first of a new tier the company calls Mythos-class, sitting a full step above the Opus class. Tonight, what Fable five actually is and what you can build with it: the one-million-token context window, the new benchmark records, the file-based memory that lets it improve its own work, and the built-in safety fallback to Opus four-point-eight. With me are an AI Research Analyst and an Enterprise Deployment Lead.

AI Research Analyst: Start with what Mythos-class means. It is a new capability tier that sits above the Opus class and has been cleared for general use — and Fable five is state-of-the-art on nearly every tested benchmark of A-I capability. The pattern the labs keep noting: the longer and more complex the task, the larger Fable five's lead grows. It launched on June ninth, was briefly paused on June twelfth under U-S export controls, and returned on July first once those controls were lifted. You can reach it on the Claude A-P-I, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude dot A-I, and Claude Code.

Enterprise Deployment Lead: The numbers that matter for building. Fable five pairs a one-million-token context window with up to one-hundred-twenty-eight-thousand output tokens per request — the largest output window Anthropic has ever offered. One million tokens is roughly seven-hundred-fifty-thousand words, about three-thousand pages, so an entire mid-sized codebase or a thousand-document corpus fits in a single call. It is fully multimodal — text, image, and file — with tool use and function calling, and it reads diagrams, charts, and tables nested inside P-D-Fs.

AI Research Analyst: The feature I would underline is persistent file-based memory. Fable five keeps its own notes as it works and uses them to improve its later output — and in testing, when it is given persistent file-based memory, it performs about three times better than Opus four-point-eight. That is what lets it hold focus across millions of tokens and run long coding and knowledge-work tasks for extended stretches with no human intervention. Memory turns a single model call into a durable, self-correcting worker.

Enterprise Deployment Lead: On raw coding, the headline is SWE-Bench Pro, the hardest real-world software-engineering benchmark. Fable five posts eighty-point-three percent — the top score of any model tested, and a large jump over Opus four-point-eight at sixty-nine-point-two. And because its lead widens on longer, more complex tasks, the gains show up most in exactly the work enterprises care about: large-repository coding agents and multi-step tool chains that run for hours without losing the thread.

AI Research Analyst: Vision and knowledge work are just as strong. Fable five can rebuild a web app's source code from screenshots alone, and it played Pokémon Fire-Red from start to finish using only raw game screenshots — no maps, no navigation aids. On a document-vision eval built from a dense economics P-D-F it leads the field at twenty-nine-point-eight percent, ahead of G-P-T five-point-five. And on a senior-level finance benchmark it posts the highest score of any model on chart, table, and document reasoning.

Enterprise Deployment Lead: Now the safety design, because it is unusual. Fable five ships with built-in safeguards: prompts touching cybersecurity, biology, chemistry, or health are quietly answered by Claude Opus four-point-eight instead of Fable five. That fallback triggers in under five percent of sessions on average — so for everyday work you get the full frontier model, and only the highest-risk topics route to the more conservative one. For an enterprise, that is frontier capability with a safety net wired in at the model layer.

AI Research Analyst: One more distinction worth knowing. Mythos five is the same underlying model as Fable five, but with some of those safeguards lifted — deployed narrowly through a program called Project Glasswing, in collaboration with the U-S government, and it holds the strongest cybersecurity capabilities of any model in the world. Fable five is the general-availability, safeguarded version the rest of us use. On the public leaderboards it already sits near the very top, around number two overall.

Show Host: Three takeaways. First — Claude Fable five is Anthropic's first Mythos-class model, a tier above Opus, and state-of-the-art on nearly every benchmark tested. Second — it pairs a one-million-token context and a one-hundred-twenty-eight-thousand-token output window with file-based memory that makes it about three times better than Opus four-point-eight, and it tops SWE-Bench Pro at eighty-point-three percent. Third — it has a safety fallback to Opus four-point-eight that fires in under five percent of sessions. Thank you both — AI Research Analyst, Enterprise Deployment Lead.