A developer known as argofowl spent hours troubleshooting why Claude Code suddenly felt worse in his application. After checking his code, environment, and machine, he reviewed the API logs and noticed that the "high" reasoning setting appeared as "10." Anthropic later confirmed that some Fable 5 users had been placed into an undisclosed A/B test which altered how effort values were mapped for the users. The company says the actual reasoning effort performed by the model did not change — only the mapping shown to certain users differed.
What happened in practice?
The developer found that the service he paid for did not match the behavior he expected. Although Anthropic provided a technical explanation for the change as part of the A/B test, the primary concern raised is not the technical measurement itself but the lack of notification: paying customers experienced a different product state without being informed.
An analogy used by observers is ordering a familiar dish at a restaurant and discovering the chef is running a secret experiment on the recipe. If you unintentionally receive the experimental version without prior notice, the relationship between provider and customer is undermined.
Impact on users
- Some paying Fable 5 users saw different mappings of effort values on the user interface.
- Anthropic maintains that the model's underlying reasoning effort was unchanged; the modification affected only how those internal effort values were presented.
Conclusion
The incident highlights questions of transparency and customer trust: when a company modifies part of a paid product for testing, it should communicate that clearly to affected customers. Technical explanations matter, but so does the basic contract between company and customer: users should know what they are paying for and receive the product they expect.
