Claude Fable 5 is the flagship you reach for on the hardest jobs, not the daily ones.

The headline capability is stamina: Claude Fable 5 can work for days at a time on a single job, planning across stages, delegating to sub-agents, and checking its own work. Most models are sprinters, great for one prompt and one answer. Fable 5 is built to be a marathon runner that holds context and intent across a long, multi-step task without drifting off course. For a small business, that is the difference between a model that helps you write one function and a model that can take on a whole build.

Anthropic calls it their most capable model for ambitious coding projects, including large migrations, complex implementations, and multi-day autonomous sessions. Four things make that possible:

• It plans, then delegates. Given a big job, it breaks the work into stages and hands pieces to sub-agents, then coordinates the results. That is closer to how a team works than how a single prompt works.
• It checks itself. Fable 5 can write its own tests to verify its work, so it is not just producing output and hoping it is right.
• It sees. It uses vision to compare its output against the goal, and it understands diagrams, charts, and tables nested inside files and PDFs, not just plain text.
• It runs in a harness. Its long-run behavior shows up when it is deployed in an agent framework like Claude Code, not in a one-off chat window. The harness is what lets it keep going.

None of this is magic, and it does not remove the need for a human to scope the work and review the result. But it does mean that a genuinely hard, multi-day task, the kind you would normally not even attempt with an AI model, is now on the table. The question is not whether Fable 5 is impressive. It is whether your automation actually needs that much horsepower.

Use Fable 5 when the job is genuinely hard and long, and Sonnet 5 would stall or need too much hand-holding to get there. That is a narrow set of situations, and that is the point. Here are the four where the flagship actually earns its price for a small team.

1. A large code migration or rebuild

Moving a legacy app to a new framework, untangling years of accumulated logic, or rebuilding a system from the ground up. These are multi-day jobs with a thousand interlocking decisions, exactly what Fable 5 was built for. It can hold the whole plan in view, work through it in stages, and test as it goes, instead of producing a plausible-looking change that breaks three things elsewhere.

2. A complex, multi-step build that has to hold together

A new internal tool, an integration that touches five systems, a data pipeline with real edge cases. When the work is too big to fit in one prompt and every step depends on the last, a model that can plan, delegate to sub-agents, and keep the pieces consistent is worth the premium. A lighter model tends to lose the thread on jobs like this.

3. A long autonomous run that plans across stages

Some tasks are not hard so much as long: process this entire backlog, reconcile a year of records, work through a large research or analysis job end to end. If it needs to run for hours or days without a human babysitting each step, Fable 5 is the model that stays coherent over that distance. Just wire in checkpoints so you can see progress.

4. Deep document or codebase reasoning with self-checking

Reading a dense contract set, a large codebase, or a stack of technical PDFs and reasoning across all of it, then verifying its own conclusions. Fable 5's vision and self-testing shine here: it can check diagrams and tables against the text and catch its own mistakes before they reach you. For a one-off summary, a cheaper model is fine. For a high-stakes read where being wrong is costly, the extra rigor pays.

The five steps at a glance: (1) make Sonnet 5 your default and Fable 5 a deliberate branch, (2) give Fable 5 a real agent harness, (3) turn on prompt caching to cut the input bill, (4) keep a human on the checkpoints of long runs, and (5) know the guardrails before you point it at sensitive data.