Best of your X follows: June 9

Today: Anthropic launches Claude Fable 5, the first publicly available Mythos-class model with built-in safety routing; Sam Altman publishes OpenAI's full third-phase roadmap; Simon Willison notes both major labs now have confidential S-1s with the SEC; Ethan Mollick surfaces a wide-ranging NYT debate on AI and jobs; and Anthropic's research team explains why bio agents lag coding agents by years.

Anthropic just released Claude Fable 5, the first time a Mythos-class model has been made available to the general public. The official announcement describes it as state-of-the-art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision. One striking data point: 80.3% on SWE-Bench Pro and 85% on OSWorld. It can run for days, and the longer the task, the larger its lead over prior models.

The key design decision is how Anthropic handled the power-safety tension. Fable 5 and Mythos 5 share the same underlying model, but Mythos 5 — so far limited to Project Glasswing partners — has certain safeguards lifted. When Fable 5's classifiers detect a request touching cybersecurity, biology/chemistry, or distillation, the request is quietly routed to Claude Opus 4.8. Users are notified every time this happens; more than 95% of sessions involve no fallback at all.

Fable 5 is available today on paid plans, in Claude Code, on the Claude API, and all major cloud platforms. It's included in paid plans at no extra cost through June 22.

Sam Altman shared a link to OpenAI's new strategic document, "Built to Benefit Everyone: Our Plan." It's the clearest public articulation of where OpenAI sees itself headed. After Phase 1 (pure research) and Phase 2 (becoming a product company), Phase 3 is about making advanced AI abundant, affordable, and accessible to every person and organization.

Three stated priorities stand out: building an automated AI researcher (internally estimated to contribute a significant fraction of OpenAI's own AI research by March 2028), accelerating the broader economy, and eventually giving every person on Earth access to a personal AGI. The document is also explicit about the risk of power concentration — calling for national and international coordination, shared safety standards, and broad distribution of AI capability rather than control by a small number of institutions.

Simon Willison flagged the symmetry: Anthropic filed its confidential S-1 on June 1st, and now both of the two largest frontier labs are in the SEC's review queue. Willison noted that Axios reported Anthropic's organic revenue growth is something the interviewer "could not find any company in any industry in any era that has scaled organic revenue this quickly at this level." The run-rate figure cited was moving from $30B to $47B.

Mollick participated in a New York Times roundtable alongside Daron Acemoglu, Dean Ball, and Clara Shih. He flagged it as "a really nice overview of the core debates" on AI and jobs, with concrete examples. The piece is behind the paywall, but the thread gives a sense of the framing — who wins, who doesn't, and what the structural choices actually are.

Mollick also flagged something subtle in the two labs' roadmap documents published this week: both mention the possibility of slowing AI development, but only under conditions of world-coordinated action using methods that don't yet exist. The framing matters — it's an acknowledgment of the question without any current operational constraint.

Anthropic published a science blog post digging into a question that's been quietly underappreciated: why are coding agents years ahead of biology agents? The answer isn't model capability — it's infrastructure. Software was built for machines: structured APIs, version control, documented packages, fast feedback loops. Biology databases were built for humans, and they show it. Heterogeneous file formats, fragmented databases, retrieval logic that lives only in web UIs, no unified machine-callable interface.

The core bottleneck Anthropic identifies isn't reasoning — it's the absence of a reliable deterministic data query layer. A scientist can articulate what they need; a biology agent often has no reliable path to actually retrieve it. This framing has direct implications for where infrastructure investment needs to go before bio agents can reach coding-agent parity.

LeCun amplified the release of VLA-JEPA in Hugging Face's LeRobot framework. What makes it different from prior vision-language-action models is that it doesn't just learn what action to take from observation — it builds a world model and predicts intermediate states before committing to a move. That's closer to the JEPA architectural philosophy LeCun has long advocated: learning internal representations of how the world works, not just mapping inputs to outputs. The technical thread from @LeRobotHF goes deep on the architecture.