You waited and Claude Fable 5 is finally here. It's Mythos-class and the most capable model we've ever GA'd, and we're inviting you to be amongst the first to build with it. We believe powerful AI can compress decades of progress into years - disease eradication, massive GDP acceleration, life extension - but AI doesn't do any of that on its own. People do. Edison once said "genius is 1% inspiration and 99% perspiration" but that ratio is inverting. The perspiration is getting automated. Humans have ideas. Agents implement, test, and evaluate them an order of magnitude faster than before. It's becoming clear that much of what advances the frontier is automatable and progress is mostly a function of tools and resources: -how fast you can run experiments -how many at once -how quickly you get answers Every important technology became important because someone picked it up early and aimed it at a problem that mattered. This Saturday, June 13th, with our friends at @cerebral_valley in SF, we're giving you a day to do exactly that. Pick up the problem that's been nagging at you and do something about it. It will be a day of building and learning with @AnthropicAI co-founder Jared Kaplan, Research PM Jeremy Hadfield, and technical leaders from @Replit, @periodiclabs, and @_hex_tech. With over $150K in Claude credits for winners, this is the day you go after: -the idea you've been sitting on -the broken workflow in your industry -the tool you wish existed The longer and more complex the task, the better. Come build something that matters and we'll see you in person on Saturday 👇

Internally, writing code stopped being the bottleneck at @AnthropicAI - our engineers ship 8x as much code per quarter as they did from 2021 to 2025. Claude Fable 5 pushes the frontier further by compressing weeks of asynchronous work into minutes and it is state-of-the-art in every benchmark: - 80.3% on SWE-Bench Pro (+11.1pp vs Opus 4.8) - 29.3% FrontierCode (+15.9pp vs Opus 4.8) - 17.4% on AutomationBench (+1.9pp vs Opus 4.8) In a nutshell, per the evals of @cognition's FrontierCode, using Fable 5 just on `low` effort  costs less and still beats the next best model, GPT 5.5 on `x-high` effort by +10%.

We've been very thoughtful about this launch: over 1,000 hours of external red-teaming produced no universal jailbreak, and every request runs past safety classifiers trained to detect misuse in cybersecurity and biology. In the rare instance a cyber or bio classifier triggers, the request gracefully falls back to Opus 4.8 instead of Fable 5, and we won't charge you Fable rates on the attempt. anthropic.com/news/claude-fa…