AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people pick a Claude model the way they pick a lift: whichever button is already lit. Others assume the dearest, newest-sounding name must be the right one for anything that matters. Both habits cost money or time, and neither tells you whether the job needed the bigger model.
The rule: start on Opus, move up to Fable only when a fair test on Opus has failed, and move up for a reason you can name.
A note on names before anything else. At time of writing, Anthropic's models overview lists Claude Fable 5.1 and Claude Opus 5.5 as the current pair, with the earlier Claude Fable 5 and Claude Opus 5 shown as "legacy models (still available)". The question "Fable 5 or Opus 5?" is therefore really a question about the two tiers. This guide uses "Fable" and "Opus" for the tiers and names a version only when a fact belongs to it. Check the overview page before you rely on any version number, because it changes.
The ladder has four rungs. You only climb when the rung below has been given a fair chance.
| Rung | What you do | Climb when |
|---|---|---|
| 1. Opus, default settings | Run the real task, with your real files and instructions | The output misses the standard you wrote down |
| 2. Opus, higher effort | Raise the effort setting (API and Claude Code expose it) and re-run the same task | Still short after a genuine second run |
| 3. Fable | Run the identical task, unchanged | Only now compare the result with rung 2 |
| 4. Fable, higher effort | Same again | Rarely needed; check the cost first |
Anthropic's own "Choosing a model" page says tuning effort is "often a better lever than switching models", and its capability-first path ends with the same advice: if evals on Opus at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work, move to Fable. The ladder is that guidance made into a habit.
Fable is described in the docs as the model "for demanding reasoning and long-horizon agentic work". Anthropic's selection matrix gives these example uses: agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck.
Reach for it when:
Fable is slower and costs more. At time of writing the API list price is $10 per million input tokens and $50 per million output tokens (about £7.50 and £38 at time of writing; check the £ price on your bill). Its cache reads are cheaper than on other models, which helps long sessions that re-read the same context.
Opus is the default starting point. The docs call it the model "for long-running agentic coding and knowledge work", and say to start with it for most workloads. At time of writing Opus 5.5 is $4 input and $20 output per million tokens (about £3 and £15 at time of writing), and Anthropic says it performs at the level of Fable 5.1 on most work at lower cost. Treat that as the vendor's claim: your own task is the test.
Use Opus for:
stop_reason: "refusal" and set up fallback.Use the Two-Run Card. It takes ten minutes and replaces a habit with a decision.
A prompt that helps you set the criteria first:
You are helping me test which AI model suits a task.
Task: [TASK_DESCRIPTION]
What a good result looks like: [SUCCESS_DESCRIPTION]
Things that must never happen: [DEAL_BREAKERS]
Do this in order:
1. Ask me for anything missing above rather than guessing.
2. Turn my success description into three pass/fail checks I can score without judgement calls.
3. Do the task.
4. Before you finish, mark each check as passed or failed, and say where you are unsure.
Format: the work first, then a short checklist table.
Fill in the task, what good looks like and the deal-breakers. Run it unchanged on each rung so the comparison is fair.
Say you run a small property-management firm and want a model to turn forty tenancy emails into a tidy issues log. You set three criteria: no invented dates, every row names its source email, duplicates merged. Opus at default passes two of three; the duplicates are missed. You raise effort and re-run: all three pass. You stop there. Fable was never needed, and you have a line for the next run: "Opus, higher effort, pass".
Now a different job: an agent that migrates a large codebase over several hours. Opus at higher effort stalls on a bug it introduced early and builds on top of it. Running Fable on the same job gets through. That is a named reason to climb, so you pay for Fable on this job and keep Opus for the rest.