Claude Fable 5

Fable 5 sits at the top of Anthropic's lineup, above Opus, and it is priced accordingly at $10 in and $50 out per million tokens. That ratio matters more than the headline: output costs five times input, so the cost of a Fable 5 task is driven almost entirely by how much it writes, not how much you give it. Feeding it a long document is cheap. Asking it to produce a long document is not.

The model is built for problems where being right matters more than being fast. It holds a thread across a long chain of reasoning without drifting, and it is noticeably better than smaller Claude tiers at noticing when a question contains a false premise rather than answering it anyway.

Two things stand out in practice. The first is long-context reliability — not just accepting a large context window, but actually using material from the middle of it, which is where most models quietly degrade. The second is prose quality: Fable 5 writes with structure and restraint, and needs less editing than anything else in the lineup, which is a real cost saving on any workflow where a human has to clean up afterwards.

It is also the most useful Claude tier for agentic work that runs many steps, because errors compound. A model that is slightly more reliable per step is dramatically more reliable over twenty.

It is expensive, and most tasks do not need it. Using Fable 5 for summarisation, classification, extraction or routine drafting is paying flagship prices for work Haiku handles. It is also slower, which makes it a poor fit for anything user-facing where latency is visible.

The honest failure mode is overuse: teams standardise on the best model, then discover the bill is dominated by tasks that never needed it.

Reach for Fable 5 when a wrong answer is expensive — legal and financial analysis, architectural decisions, high-stakes drafting, or agent runs long enough that per-step reliability compounds. For everything else, start at Sonnet 5 and move up only when you can point at a specific failure. The right pattern is routing, not standardising.