Stop telling Fable what to do
My team ran every Claude model and effort level across hundreds of tasks. Here is what we actually learned about which model to use, and how to talk to it.
There are a lot of Claude models and effort levels now. Over the last few weeks my team has run all of them across hundreds of real tasks, so I want to share what we learned about when each one earns its keep. I will give an honorable mention to ChatGPT 5.5 as well. We will not be talking about Gemini, because right now I do not think you should be using Gemini for work.
First, the clock. Anthropic just announced it is extending the subscription deadline for Fable 5 from July 7 to July 12. After that, Fable is usage-priced only, which makes it prohibitively expensive for most of us. Unless they extend again. Please, please, please, Anthropic. I am already grieving.
The Fable skeptics are holding it wrong
While you still have it, use Fable 5 for any task that is both difficult and important. And if you are one of the people saying Fable is no better than Opus, you're wrong. I'm sorry, you're just wrong. My best guess is that you are not respecting how smart it is, and you are treating it like a dumber model. Dumber models needed to be told not just what to do but exactly how to do it, step by rigid step, with reminders, or they went off the rails. Since around Opus 4.6 you could just say what you wanted and let the model find the steps, which is the whole agentic shift in miniature. Fable is another paradigm again. Stop telling it what to do. Tell it the goal. If you want a dashboard, do not specify what goes on the dashboard. Tell it what the dashboard is for and what it needs to accomplish. Fable 5 will usually make better calls than you would have.
Effort level is simpler than people make it: run Fable on high, full stop. Below high you get a dumbed-down Fable. Above high you are mostly setting tokens on fire. Fable extra, max and ultracode do not summon a smarter model. They spend more tokens on the same one, run more checks, spin up more adversarial agents. Ninety five percent of the time you get the same answer high would have given you. Unless you work at Anthropic and tokens are free, high is the setting.
Opus is the opposite case
If you do not have Fable, Opus is the best model for most day-to-day work. And here the effort math flips. On difficult, important tasks, Opus extra, max and even ultracode genuinely pay off, because Opus simply makes more mistakes than Fable, and double-checking work that contains mistakes is worth the tokens. Just be honest with yourself about whether the task justifies the burn. For everything routine, Opus on high is a great default.
ChatGPT 5.5 inside Codex deserves its mention: for some jobs it is the best tool, and on a Codex subscription it is absurdly cheap. A trick we use daily: tell Claude that whenever a task calls for computer use, image generation, or deep research across the web, it should call the Codex CLI and hand the work over. 5.5 is also a sane fallback when you are low on tokens and the task is neither difficult nor important.
Which leaves Sonnet. Sonnet is the cheap model for genuinely easy work. Drafting an email, reformatting a document, fine. Hand Sonnet a hard problem and it will thrash through so many tokens that you end up paying Opus prices for Sonnet answers. Match the model to the job and the job to the goal, the same way you would match a platform to a workload. Then get out of the way.
Opinion columns reflect the personal views of the author. Our reporting on the stories referenced here lives on the linked pages.