Anthropic has just announced Fable, the start of a new family of models that brings many of the capabilities of its Mythos system to the public. As a refresher, Mythos is the state-of-the-art model Anthropic debuted at the start of April through Project Glasswing. The project saw Anthropic share access to the model with select partners, including Apple and NVIDIA, with the aim of helping those organizations harden their software against AI cyberattacks. Glasswing also prompted the White House to rethink its policy on AI regulation. Last week, Anthropic said it was working as quickly as possible to release a Mythos-capable model with “highly robust safeguards” to the public, and today it’s doing exactly that with Fable.
Anthropic says the new system, which starts at version five to align with the company’s other models, offers capabilities that “exceed” those of any it has made publicly available in the past. “The longer and more complex the task, the larger Fable 5’s lead over our other models,” Anthropic says. In the company’s own testing, Fable beat not only its previous flagship offering, Opus 4.8, but also GPT-5.5 and Gemini 3.1 Pro from rivals OpenAI and Google. Anthropic adds Fable is particularly adept at tasks like software engineering, drug design and document analysis, but also excels in domains where the company’s past models have typically underperformed. Vision is one such area.
https://www.youtube.com/watch?v=CIQBP1w4B1M
You may recall last year Anthropic streamed Claude trying (and failing) to beat Pokémon Red. The model Anthropic used for that Twitch series, 3.7 Sonnet, required an overlay to allow Claude to make sense of the game’s pixel art interface and world. If you watched the stream closely, you would have noticed each time Claude moved the game’s protagonist it had to reevaluate their position afterward. Between frames, Claude was essentially blind, and it couldn’t “hear” when the player character hit an obstacle like a tree. By contrast, Fable can beat FireRed, the 2004 Game Boy Advance remake, with a “minimal, vision-only harness.” In practice, the company says Fable’s new capabilities translate to a model that can “extract precise numbers from detailed scientific figures and can perform complex vision-based tasks like rebuilding a web app’s source code from screenshots alone.”
As for those safeguards I mentioned earlier, Claude will automatically route prompts on some topics through the less-capable Opus 4.8. “To release [Fable 5] both safely and quickly, we’ve tuned these safeguards conservatively — they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5 percent of sessions,” the company said. “With more capable models arriving in the coming months, we’re working to improve our safeguards and reduce false positives as quickly as we can.”
On Tuesday, Anthropic also released Mythos 5, which runs on the same underlying model as Fable 5 but comes with fewer safeguards. The company will initially offer Mythos 5 through Project Glasswing before making it more widely available through a trusted access program. Claude subscribers can try Fable 5 until June 22. Thereafter, using Fable 5 will cost usage credits, which Anthropic has priced at $10 per million input tokens and $50 per million output tokens. “After this point — when sufficient capacity allows us to do so — we aim to restore Fable 5 as a standard part of subscription plans. We intend to do this as quickly as we can.”