Anthropic says AI built a prototype operating system
The new models cost less, but Anthropic still recommends Opus for most everyday coding and knowledge work.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:005:03
Read transcript
Anthropic says its new Fable model coordinated more than a thousand AI subagents to build a prototype operating system from an empty folder. The eight-day experiment produced a system that boots on a laptop, with drivers, file storage, windows, a login screen and an installer. It’s a striking demonstration of long-running, delegated coding—but it’s an experiment, not evidence that the model can reliably build production operating systems. Anthropic announced Fable 5.2 and restricted-access Mythos 5.2 for demanding engineering work. The two share an underlying model, but differ in safeguards and who can use them. Fable is generally available, with protections Anthropic says block most cybersecurity-related tasks. Mythos is being offered to selected organizations through cybersecurity and life-sciences verification programs, with fewer safeguards in those fields. Anthropic also describes a separate production project: agents using an earlier Fable version migrated a service from Python to Rust, and the company says startup became about 26 times faster. That’s a company-reported result, not an independent evaluation. The same caveat applies to its benchmarks. In early FrontierCode 2.0 testing, Fable at High effort scored 45.3 percent, close to Opus 5.5’s best score of 45.5 percent, at roughly half the cost per task. The benchmark asks whether maintainers would accept changes to open-source code; it doesn’t settle how a model will perform across real engineering work. Both new models cost eight dollars per million input tokens and 40 dollars per million output tokens. Optional fast mode doubles those prices, with Anthropic saying generation can be up to 2.5 times faster. Despite the lower costs, the company still recommends starting with Opus 5.5 for most coding and knowledge work, and switching to Fable or Mythos when the assignment calls for it. The distinction is practical: the newest or cheapest model isn’t automatically the right default. That distinction between a general-purpose default and a specialist also shows up in Meta’s new model for routine agent work. Muse Ember is designed for the repeated steps between bigger decisions: calling tools, reading documents, running tests and making targeted code fixes. Meta’s proposed pairing leaves harder planning to Muse Wonder, with Ember handling the volume of smaller tasks. Ember is available through Muse Code and the Meta Model API, starting at 10 cents per million input tokens and 50 cents per million output tokens. Meta says it can process text, images, audio and video, which could make it useful for document-heavy workflows as well as coding. The launch results support a focused, rather than all-purpose, pitch. Meta reports a strong showing on a business-workflow benchmark, while coding results are mixed: Ember beat some models on one test, but trailed Claude Haiku on a benchmark measuring whether real engineering changes could be merged. These are company-presented test results, not a guarantee for any particular workload. For builders, the useful question is whether a low-cost delegated step is good enough—and whether the cost of checking its work stays low too. The same effort to make AI tools more immediately usable is behind Google’s Maps integration in AI Studio. A developer can describe a location-based app in a prompt, and Gemini can build it while the workspace handles Maps configuration and credentials, then shows an interactive preview. Google’s examples include store locators, route planners and 3D city tours. The no-credit-card sandbox allows 100 calls per day for each API, with quotas resetting every 24 hours. That’s a way to prototype, not unlimited free production use. Apps that need more capacity or are ready to launch require a linked Cloud billing account and follow standard Maps Platform pricing, including recurring monthly free usage. Google says the setup can turn an idea into an app in minutes; that speed is its claim, not an independent measurement. For teams trying an idea quickly, the practical boundary to watch is when a working preview becomes a real service: that’s when API quotas and billing move from setup details into product decisions. And as AI moves into everyday work, Medium is making the writing process itself more visible. Its free web app, TK, puts an assistant inside the editor: writers can call it with “at TK” to develop an idea, retrieve saved writing or get feedback on a draft. They can also publish notes describing how AI contributed, and share early drafts with people for human feedback. TK is available to anyone with a Medium account and supports Markdown import and export. Writers can optionally connect their own AI agent, but don’t need one to use the app. One important limitation: writing in TK is not encrypted. Medium says users can set access permissions, but those permissions don’t change the encryption status. A desktop app is planned; for now, TK runs in a browser. Across these launches, access is getting easier, but the real decision is still about boundaries: which tasks can be delegated, what safeguards and review they need, and what changes when a prototype becomes production. Watch the total cost of checking AI work—not just the price per token or the speed of the first demo.
