# Fable 5 Pushes Coding Agents Into Project-Sized Work

*By Coding Agents Alpha Tracker • June 10, 2026*

Claude Fable 5 dominated the day, but the useful signal is practical: senior engineers are handing coding agents bigger objectives, longer autonomous runs, and cleaner human-approval seams. This brief distills the highest-leverage workflows, what actually shipped, and the clips and code worth studying.

## 🔥 TOP SIGNAL

- Claude Fable 5 is the clearest step-function signal today: Andrej Karpathy says the jump shows up on long, difficult problem-solving sessions because you can hand the model more ambitious end-to-end tasks and it will run with them; Mike Krieger says it is the first model he hands whole projects to; Simon Willison used it to ship major agent features in [LLM 0.32a3](https://llm.datasette.io/en/latest/changelog.html#a3-2026-06-09) and a 13.9 MB CPython WASM wheel in a day. The practical takeaway is larger task envelopes and longer autonomous runs, not blind trust: Karpathy warns against fully hands-off production use because launch safeguards are still trigger-happy. [^1][^2][^3]

## ⚡ TRY THIS

- **Shift from tasks to objectives in Fable.** Alex Albert's launch checklist: 1) give it bigger tasks, 2) define what done looks like and how to verify it, 3) use `/loop` and `/goal` for open-ended work, and 4) rework old `CLAUDE.mds` or skills that may anchor it to older-model behavior. Karpathy's examples of the bigger task class: explainers, visualizers, dashboards, bespoke single-use apps, 10x test suites, auto-optimization, and custom research HTML. Albert recommends `xhigh/high` by default, while Matthew Berman found the lowest effort setting sufficient in many of his tests and called extra effort slow or overkill. [^4][^1][^5]

- **Use repo + artifacts + iterative follow-ups instead of one giant prompt.** Simon Willison started with `Clone simonw/micropython-wasm from GitHub and research how this could use a full Python as opposed to MicroPython`, attached Brett Cannon's build zips, then followed up with `Try a bit more at the single-zip-stdlib problem` and `I want a wheel that has the whole system in it...`. In his run, that produced a 13.9 MB wheel usable via `uv run --with <wheel> cpython-wasm -c 'code'`. [^3]

- **Add a pause/resume seam for human approval.** Simon used Fable to push a clean pattern into `llm`: let tools accept `llm_tool_call`, raise `llm.PauseChain` when you need approval, and resume from `messages=` history or explicit `tool_calls_list=` so the chain keeps tool-call state. This came out of Datasette Agent's `ask_user()` requirement and is the best public example in today's source set of human-in-the-loop tool execution without brittle placeholder hacks. [^3]

- **Run a pre-prod repo review, then override the question treadmill if needed.** Berman's test prompt is `review the entire code base` and ask for security, docs, logical gaps, edge cases, UX/UI workflows, and test coverage; in workflows mode he saw that fan out into 100+ subagents. swyx's lighter-weight version is `review my code for issues` as a pre-prod `Fable Check`, and Berman's override when Claude gets stuck in clarification/spec-confirmation loops is `Don't ask more questions. Just build.` [^6][^7][^5]

## 📡 WHAT SHIPPED

- **Claude Fable 5** — Mythos-class model with added safeguards; Anthropic says its lead grows as tasks get longer and more complex. Shipping today in Claude Code/Cowork, paid Claude plans, the API at `$10/M` input and `$50/M` output, and Devin/Cognition at `1.4x ACUs`; cyber/bio requests fall back to Opus 4.8 and Mike Krieger says `95%+` of sessions never notice. Public benchmark numbers cited across today's coverage: SW Bench Pro `80%`, Frontier Code Diamond `29.3%`, and computer use `85%`. [^1][^8][^9][^10][^11][^6]

- **Cursor integration** — Fable 5 is live in Cursor and set a new CursorBench SOTA at `72.9%`, `8` points above the previous best; comparison page is [cursor.com/evals](http://cursor.com/evals). Practical warning from Jediah Katz: Cursor users should read the model docs first because Fable has unusual required data-retention policies for safety. [^12][^13][^14]

- **Claude Code nested subagents** — shipped today with depth capped at `5`. [^15]

- **[LLM 0.32a3](https://llm.datasette.io/en/latest/changelog.html#a3-2026-06-09) / [Datasette Agent](https://github.com/datasette/datasette-agent)** — Fable pushed out `llm.PauseChain`, `llm_tool_call`, guaranteed unique `tool_call_id`, and resuming from unresolved tool calls; Simon says the release was almost entirely written by Fable. Adoption signal: he spent `$110.42` of tokens in one day on his Max plan, with one `prod_datasette_agent` session accounting for `78.2M` tokens and `$99.26`. [^3]

- **Anecdotal side-project ceiling** — Riley Brown says Fable one-shotted a Replit-like mobile app builder from the prompt `build an app like Replit, that uses @daytonaio for sandboxing, and convex for DB`, and later extended a Lovable-like builder to web/mobile preview, editing, sandboxes, DB/auth, voice-to-text, and Swift previews in `2-5` prompts. Treat this as anecdotal, but it is a useful snapshot of what fast-moving builders tried on day one. [^16][^17][^18][^19]

## 🎬 GO DEEPER

- **Matthew Berman — 13:05-13:31: workflows fan-out.** Best short demo of what repo-wide orchestration looks like in practice: a planning agent that can spawn 100+ subagents after a few minutes of planning. Good watch before you hand a big review or migration to workflows mode. [^6]

[![MYTHOS MYTHOS MYTHOS](https://img.youtube.com/vi/Ou-0vjl6FZo/hqdefault.jpg)](https://youtube.com/watch?v=Ou-0vjl6FZo&t=784)
*MYTHOS MYTHOS MYTHOS (13:04)*


- **Matthew Berman — 23:36-24:22: model routing for cost discipline.** His takeaway after a week with Fable: reserve it for the hardest long-horizon problems and route the rest to Sonnet/Haiku. [^5]

[![MYTHOS is LIVE!!!!](https://img.youtube.com/vi/e-4xG1U7M70/hqdefault.jpg)](https://youtube.com/watch?v=e-4xG1U7M70&t=1415)
*MYTHOS is LIVE!!!! (23:35)*


- **Matthew Berman — 48:51-50:12: the `Just build` override.** Worth watching if you keep hitting clarification, spec confirmation, and approach approval loops before the model starts coding. [^5]

[![MYTHOS is LIVE!!!!](https://img.youtube.com/vi/e-4xG1U7M70/hqdefault.jpg)](https://youtube.com/watch?v=e-4xG1U7M70&t=2931)
*MYTHOS is LIVE!!!! (48:51)*


- **Read these diffs:** [Datasette Agent PR #20](https://github.com/datasette/datasette-agent/pull/20) and [LLM PR #1482](https://github.com/simonw/llm/pull/1482). They show a real human-approval workflow getting turned from local hacks into framework primitives. [^3]

- **Tool worth adding:** [AgentsView](https://agentsview.io/) for local agent transcript and cost analysis. Simon uses it to attribute spend across projects and models; when Fable pricing was missing, he used Fable itself to reverse-engineer a custom price config. [^20]

*Editorial take: the bottleneck just moved up a level — from getting a model to write code, to deciding how much work to hand it, where to force approval, and when frontier-model tokens are actually worth spending.* [^1][^4][^3][^5]

---

### Sources

[^1]: [𝕏 post by @karpathy](https://x.com/karpathy/status/2064409694761054332)
[^2]: [𝕏 post by @mikeyk](https://x.com/mikeyk/status/2064392825480032418)
[^3]: [Initial impressions of Claude Fable 5](https://simonwillison.net/2026/Jun/9/claude-fable-5)
[^4]: [𝕏 post by @alexalbert__](https://x.com/alexalbert__/status/2064467657483829441)
[^5]: [MYTHOS is LIVE!!!!](https://www.youtube.com/watch?v=e-4xG1U7M70)
[^6]: [MYTHOS MYTHOS MYTHOS](https://www.youtube.com/watch?v=Ou-0vjl6FZo)
[^7]: [𝕏 post by @swyx](https://x.com/swyx/status/2064492823781789969)
[^8]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2064394151441863006)
[^9]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2064402671898075579)
[^10]: [𝕏 post by @mikeyk](https://x.com/mikeyk/status/2064392996288901392)
[^11]: [𝕏 post by @swyx](https://x.com/swyx/status/2064396531231510931)
[^12]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2064394824313376787)
[^13]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2064394826100134193)
[^14]: [𝕏 post by @jediahkatz](https://x.com/jediahkatz/status/2064394895414948038)
[^15]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2064327225504403752)
[^16]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2064515211894091869)
[^17]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2064540450317271356)
[^18]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2064568569061331065)
[^19]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2064560958660300962)
[^20]: [Setting a custom price for a model in AgentsView](https://simonwillison.net/2026/Jun/9/agentsview-custom-model-price)