Anthropic's Tariq Shihipar delivers a four-part field guide to Fable, a new "mythos" class of model rolling out that afternoon: unhobble Claude to unlock its capability overhang, use Fable to find your own unknown unknowns, work through the grief of no longer hand-coding, and be unreasonable because trade-offs are no longer real. His resolution: do the best work of your life faster than ever, then work less.
Opening the Day 2 keynotes, Anthropic's Tariq Shihipar (Claude Code) announced that Fable — a model teased the day before — was rolling out later that afternoon, with more detail promised in a 12:30 fireside chat alongside Cat Woo and Simon Willison. He described it as one of those Anthropic models people just remember, in the lineage of Sonnet 3.5 (new), Opus 4, and Opus 4.5.
His metaphor: Fable is like reaching the end of an RPG tutorial and watching the open world unlock — thrilling but also intimidating, because so much becomes possible. To help, he offered a "field guide" with four parts: unhobbling Claude, finding your unknowns, dealing with the grief, and being unreasonable.
Shihipar argued that models are grown, not designed — Anthropic gives them data, feedback, and compute, then learns their behavior empirically. Because of that, what constrains a model is us: the harness we put it in and the way we prompt it are a function of our understanding. "Unhobbling" means understanding Claude well enough to unleash it, and there's far more to unlock in Fable.
He illustrated capability overhang with a viral question: which Pokemon end in "aw"? A plain chat model can't answer despite knowing every name, but Claude Code fetches all the Pokemon and writes a script to filter, correctly returning two. Claude gets smarter in spiky ways — give it the right tool and new capability appears. He traced this progression through pasting context vs. giving Claude the bash tool to build its own context (the insight behind Claude Code), proactive multiplayer work (Claude tag), the ask-user-question tool growing from barely usable on Opus 4 to full HTML reports on Opus 4.8 and Fable, and the shift from markdown to rich HTML reports.
A concrete sign of the new class of models: Anthropic recently removed 80% of Claude Code's system prompt. Best practice used to be a small prompt with few tools and many examples (Sonnet 3.5 new era), then larger prompts with many examples and tools as models improved. Now the newest models want a smaller prompt — examples tend to constrain a model that is more imaginative than the examples, so Anthropic favors giving context over constraints and avoids "do not do this" instructions.
Shihipar emphasized this is closer to biology than physics: empirical, organic, without fully known rules but with intuition to build. He recommended Anthropic's paper on the biology of a large language model as a way to build that intuition.
Working with Fable, Shihipar said, you also have to unhobble yourself. "The map is not the territory": your plan, prompt, and spec are the map, but the codebase and real-world constraints are the territory. Wherever Claude hits something in the territory that isn't on the map, that's an unknown — a decision point you didn't specify. Fable traverses so much ground that unspecified unknowns multiply, so he uses a known/unknown matrix (known knowns, known unknowns, unknown knowns, unknown unknowns) to reason about them.
He shared techniques for using Fable to find its own unknowns: a blind-spot pass where Claude scans a module, git diff, or Slack for gotchas; brainstorm prototypes (e.g., an HTML page with four wildly different designs) to surface "know it when I see it" preferences; having Claude interview you, prioritizing questions that would change the architecture; passing reference code as another map instead of a written spec; asking Fable to log implementation notes when it hits an unknown; and finally quizzing you afterward so you stay in the loop and can represent the work in a PR.
Shihipar spoke candidly about loss. Reflecting on a YC startup he once ran with about 30 people, he recalled constant trade-offs because code was so hard — features that took a month or two. Returning to that codebase recently, work that would have taken weeks took hours. He genuinely loved writing code by hand, but also remembers "swimming in failure" through late nights and failed projects, and concludes he can't go back: "the only way out is through."
The fourth part is being unreasonable. He credits Anthropic's culture with the belief that trade-offs aren't real — instead of prioritizing one thing against another, ask what if you just did all of it and forced reality to show you the trade-off. Good/fast/cheap becomes "pick three." He made his talk's deck in about four hours with Fable, and framed the mission: the world is looking to AI engineers to prove AI works, so do the best work of your lives faster than ever, then work less and spend more time with people you care about. Building is easier now, but generating value is still hard — and that, not the setup, is the goal.
the models are grown not designed
the map is not the territory
we believe that trade-offs are not real
the only way to prove that agents work is to do the best work of our lives faster than ever before
| Time | Topic |
|---|---|
| 00:00 | Intro, selfie, and Fable rolling out today |
| 01:34 | The field guide: four parts |
| 02:05 | Unhobbling Claude & capability overhang (the Pokemon example) |
| 05:44 | Removing 80% of the system prompt; new tools and HTML reports |
| 08:52 | Finding your unknowns: the map is not the territory |
| 14:08 | Dealing with the grief of leaving hand-coding behind |
| 16:14 | Being unreasonable: trade-offs are not real |