Start with Opus 5 for most work. Try Fable 5.1 on a defined demanding task, and widen its use only when comparable results justify it. This is provisional advice based on documentation, not a hands-on verdict.
I have not recorded hands-on Fable 5.1 results. API prices do not establish subscription value or total cost per completed task. App defaults, availability and model switching can differ from API behaviour. Sources checked on 7 September 2026.
Repeatable results on representative tasks showing that Fable 5.1 gives a better usable result for the total cost and checking required. One successful task would justify that use case; a wider recommendation needs broader evidence. A changed default alone would not establish better results.
Start with Opus 5 for most work, and give Fable 5.1 a specific job to justify a switch. That’s a provisional rule based on Anthropic’s documentation.
What launched on 1 September
Anthropic’s Fable page dates the Fable 5.1 launch to 1 September 2026 and lists access for Pro, Max, Team and Enterprise. API prices are $10 per million input tokens and $50 per million output tokens. Cache reads, which reuse previously processed input, cost $0.25 per million tokens, 75% below Fable 5. Anthropic estimates savings of about 25% on typical workloads and up to about 45% on highly agentic workloads. These are vendor estimates, checked on 7 September 2026.
The model overview lists Opus 5 at $5 input and $25 output per million tokens, with both models supporting a one-million-token context window. Fable’s listed input and output rates are twice Opus’s; that does not mean every completed task costs twice as much. Token use, cached input, retries and tool use affect the total. These API rates also do not tell a Pro or Max subscriber how much useful work their subscription will cover.
Anthropic's own start-here advice
Anthropic’s Fable 5.1 guidance recommends Opus 5 as the starting point for most workloads. It points to Fable 5.1 for hard reasoning and long tasks involving multiple tool actions, or when checks on Opus at a higher reasoning effort still fall short. That is the vendor’s advice, not evidence from my own comparison.
The model selected in the Claude app may differ from the starting point recommended in API documentation. Check the picker rather than assuming either model is the default.
Check which model answered
Anthropic’s model-switching help says automated safety checks can switch some cyber or biology requests to Opus. With automatic switching enabled, the picker stays there unless you change it. A switch before output is billed at Opus rates; a switch partway through charges the input and output already produced at Fable rates, then the remainder at Opus rates. The API requires configured fallbacks rather than enabling this switching by default. Record any switch when comparing results, because a chat started on Fable may finish on another model.
Try it on one defined task
For ordinary writing, planning and checking, keep the starting point you already have unless there is a problem to solve. If Opus 5 misses something that matters, first check whether the prompt and source material were sufficient. Then try Fable 5.1 on that same task, with the same inputs and clear criteria for an acceptable result.
For example, compare a difficult document review against a list of known issues. Record which issues each model found, which claims you had to correct, and whether the final output was usable. For API work, compare the actual total cost. For subscription use, compare the result and limits you encounter.
When this would change
A repeatable improvement on one hard task would justify using Fable 5.1 for that task. A wider switch needs comparable evidence across the work you do regularly, including the total cost and checking required. A new default in the picker would change what you see first; it would not, by itself, show that the model is better for your work.