Ninety percent of the code merged into Anthropic’s own codebase is now written by an AI model. Not reviewed by one. Written by one.
That number comes straight from Anthropic’s announcement of Claude Fable 5.2 and Claude Mythos 5.2, published this week. It’s a striking statistic. And it matters for anyone who builds software, runs a software team, or pays for software to be built.
Fable and Mythos are what the industry calls “frontier models,” meaning they sit at the highest capability tier available to outside developers. The two share the same underlying architecture. Fable includes safety guardrails that make it broadly available for general business use. Mythos is a more capable variant restricted to vetted organizations doing cybersecurity or biomedical research.
The improvements in 5.2 are concrete. Output generation runs up to 30% faster than the previous version, and pricing dropped to $8 per million input tokens and $40 per million output tokens. For teams running AI-heavy workflows at any volume, that cost reduction changes what’s worth automating.
On benchmarks, Fable 5.2 outperforms GPT-6 Astra from OpenAI and Gemini 4 Carbon from Google on agentic coding tasks. These tests measure whether a model can complete multi-step software engineering work without constant human direction. It scored highest of any released model on the FrontierCode benchmark at maximum effort. That’s not a marginal difference.
The more interesting story is what Anthropic did internally with an earlier Fable 5.2 build. One engineer gave the model a single assignment: design a prototype operating system from scratch. The model coordinated over 1,000 AI subagents working continuously for eight days. The result was a working operating system that boots on a laptop, complete with disk drivers, file storage, a login screen, and an installer.
Multi-agent orchestration is a production pattern now, not a research experiment. Fable 5.2 is explicitly designed to act as an “Orchestrator,” coordinating fleets of smaller AI models by breaking large tasks into pieces and delegating them. Anthropic uses it to run large-scale automation internally, and so do many enterprise teams.
For businesses that commission custom software or maintain development teams, this release deserves attention. It doesn’t mean developers disappear. It means the capable ones can compress a six-month project into weeks. The cost calculation for building software is changing, and organizations that adjust their expectations accordingly will plan better and spend smarter.
Fable 5.2 also improved on computer use, meaning it can operate a browser or desktop application to complete tasks the way a person would. Vision capabilities got a significant upgrade too, with sharper interpretation of charts, diagrams, and visual documents.
The speed increase and price drop together lower the barrier for real production use. This isn’t a research preview. It’s infrastructure.
Want to explore how the latest AI models could benefit your business? Let’s talk.