IndyDevDan: Claude Fable 5 Is an Orchestrator, Not an Intern

The 15-sandbox test

To compare the models on more than vibes, IndyDevDan had Fable 5 orchestrate three models — itself, Opus, and Sonnet — each running the same five specs in its own agent sandbox, looping until the work was complete and shipping a live public URL. That's 15 full-stack apps total (an LLM price-index clone, a Hacker News clone, a scikit-learn model generator, a multi-agent chat room, and more).

The receipts, as reported in the video:

• Sonnet: ~$55 of tokens, ~700K tokens.
• Opus: ~$91 of tokens, ~700K tokens.
• Fable 5: ~$200 of tokens, ~1M tokens.

On raw price-per-token, Fable loses — it costs more and uses more tokens. But it completed the same work roughly 20% faster. On about 80% of these tasks, Opus and Sonnet did the job at a fraction of the price; the premium only paid off on the genuinely hard one.

Observation 1 — Buy intelligent agent hours, not tokens

IndyDevDan's headline reframe: with a model like Fable 5 you're not buying tokens, you're buying intelligent agent hours — and time is the scarce resource. The premium scales directly with the difficulty of the mission. His blunt rule of thumb: "If you can cure cancer with a million Fable tokens, $10/million is nothing. If you're centering a div, you're just making a donation to Anthropic." The harder the spec, the more Fable makes sense; for trivial work, Opus or Sonnet are the rational choice.

Practically, that means getting out of a prompt-by-prompt, babysit-the-agent loop and into specs, proper delegation, and closed-loop structures where the agent can validate its own work and you judge the result.

Observation 2 — Fable 5 is an orchestrator, not an intern

The standout for IndyDevDan after using the model since release: Fable 5 isn't a worker you assign a single task to — it's best pointed at the most valuable thing an agent can do, which is orchestrate other agents. He frames an engineer's progression as: base agent → prompt/context engineering → more agents → specialized agents → orchestrating all of the above.

He points to the model's system card, which has a dedicated multi-agent section: benchmarks that battle-test the best model alone versus the same model scaled to 3, 5, 10 — and unlimited — async sub-agents. The result is what you'd expect: scale your compute and you scale your impact, provided you have a strong model steering the sub-agents. Cloudflare's "software factory" gets the same nod as a real-world multi-agent example.

Key Takeaways

• Fable 5 and Mythos 5 were pulled from subscription plans (ProMax, Team) — the video reports a federal export-control order after a jailbreak, and says re-release is expected. Treat as developing, unconfirmed reporting.
• The honest metric for a frontier model is price per intelligent agent hour, not price per token. Fable cost ~2× Opus and used more tokens, but finished ~20% faster.
• On ~80% of tasks in the test, Opus or Sonnet did the job for far less — reserve the top tier for genuinely hard, high-value missions.
• Fable 5's real edge is as an orchestrator of sub-agents, not a single-task worker. The model's system card backs multi-agent scaling.
• The workflow shift: write strong specs, delegate, build closed loops where the agent validates its own work, then judge the output — rather than prompting back and forth.