One thing is certain. Many people say Claude Fable 5.2 burns through usage fast after they try it. These are user reports, and they are not verified yet. There is no side-by-side test with the same task at the same effort level. According to Anthropic's official docs, Fable 5.2's default effort level in Claude Code is medium. If your session runs at max, a big part of the higher usage comes from that effort level.
What Users Say
These are all personal reports. The tasks, context and effort levels differ, and they are not verified yet:
- Someone used it for about twenty minutes and lost most of their usage.
- Someone only talked through a plan, with no code writing, and still hit the limit.
- Someone used it for web projects and animation. The results surprised people around them, and it used a lot of usage.
- Others feel it is not much better than the previous model tier on their own work. They keep a cheaper model for normal projects and switch to Fable only for hard problems.
One detail is worth noting: a user later found that their test session had stayed on max the whole time. Comparing the cost of top-tier Fable with a standard-tier model really compares effort levels.
Rule Out These Three Things First
| Check | How to check | Official basis |
|---|---|---|
| Effort level | /effort status |
Default is medium. The CLAUDE_CODE_EFFORT_LEVEL environment variable overrides the choice in the interface |
| Context size | /usage |
Every message carries the full history. In a long session, even a short message is billed as the whole history |
| Did you switch models mid-session? | Cache hits in /usage |
After a model switch, the whole history has to be processed again |
For how to check the effort level, see this news post. For why switching models costs more, see this one.
Run Your Own Test
To find out whether Fable is worth it, test it on your own work instead of relying on other people's screenshots. Send this to your AI agent:
I want to compare what two models cost for the same job. Please help me design a test:
1. Pick a small task from this project that takes under twenty minutes and that tests can judge as right or wrong. Write it as a fixed task description.
2. Start both runs from a brand-new session, with the same effort level, and do not change the task description by a single word.
3. After each run, record: how many tokens were used, how long it took, whether the tests passed, and how many times I had to follow up.
4. Put the results in a comparison table. State only the numbers, and do not draw a conclusion for me.
Notes
- Work like web pages and animation produces very long output, with many output tokens. It costs more than changing a few lines of code, as expected.
- Lower effort levels are faster and cheaper. Simple tasks do not need a high effort level.
- One comparison only shows results for that one task. For a different kind of work, run the comparison again.
Official details are in Claude Code's model configuration.


Comments (0)
No comments yet. Start the conversation.