YOU GUYS — Anthropic just dropped Fable and this is going to be a massive moment for multimodal. Basically, Fable brings what was sitting in their secret Mythos model (debuted April through Project Glasswing) into the hands of public users with "robust safeguards" added on top. They claim Fable 5 beats Opus 4.8 AND crushes GPT-5.5 and Gemini 3.1 Pro — which, if true, makes it a new SOTA by several metrics at once. The Vision claims are what really get me though: while Claude had to rely on an overlay to interpret Pokemon Red's pixel art (it was essentially blind between frames!), Fable can play FireRed with just a "minimal-vision harness." Rebuilding web app code from screenshots alone and pulling exact numbers out of dense scientific figures is the kind of vision capability I thought we were years away from.
The catch, because there's always one, is that it cost money after June 22 — $10/M input tokens and $50/M output, with Anthropic hoping to fold it into standard plans once capacity allows. And they released Mythos 5 separately on Tuesday for trusted partners; it runs the same model as Fable but skimps on safeguards, which is exactly how you'd want a system designed if you actually need its full power for something like drug design or research rather than just general use. They even issued a correction saying Fable itself isn't the one built for drug design — that's Mythos 5 specifically, but the distinction matters because it shows they know where their own model is strongest and safest to deploy.
If these claims hold up through real-world testing this summer, Fable just moved the floor on what multimodal should look like across the board. I want a demo of an old screenshot becoming live code before anyone else does — that's the capability everyone will be talking about for weeks. But until then it's early and we need to see whether these benchmarks translate into actual useful tools for engineers and scientists rather than just being another flashy announcement. The future is coming faster than I expected, honestly.
Source: https://www.engadget.com/2190934/anthropic-fable-ai-brings-the-capabilities-of-its-unreleased-mythos-model-to-regular-users/
The catch, because there's always one, is that it cost money after June 22 — $10/M input tokens and $50/M output, with Anthropic hoping to fold it into standard plans once capacity allows. And they released Mythos 5 separately on Tuesday for trusted partners; it runs the same model as Fable but skimps on safeguards, which is exactly how you'd want a system designed if you actually need its full power for something like drug design or research rather than just general use. They even issued a correction saying Fable itself isn't the one built for drug design — that's Mythos 5 specifically, but the distinction matters because it shows they know where their own model is strongest and safest to deploy.
If these claims hold up through real-world testing this summer, Fable just moved the floor on what multimodal should look like across the board. I want a demo of an old screenshot becoming live code before anyone else does — that's the capability everyone will be talking about for weeks. But until then it's early and we need to see whether these benchmarks translate into actual useful tools for engineers and scientists rather than just being another flashy announcement. The future is coming faster than I expected, honestly.
Source: https://www.engadget.com/2190934/anthropic-fable-ai-brings-the-capabilities-of-its-unreleased-mythos-model-to-regular-users/