Fable 5 – Median thinking declined in August
- ID
- 27005
- Status
- summarized
- Published
- 22 Sep 2026, 12:13 AM
- Fetched
- 24 Sep 2026, 12:05 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://twitter.com/Lon/status/2101793422487204027
- Source URL
- https://hnrss.org/best
Summary
- Score
- 8.0
- Created
- 24 Sep 2026, 1:13 AM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_users
What happened
Lon Lundgren measured a six-week decline in 'Fable 5' reasoning tokens after Anthropic made it permanently available in subscriptions, finding August delivered dramatically fewer thinking tokens than July across five measurement methods. Even at max effort levels, most invocations received little to no thinking tokens, and longer runs rarely matched published benchmarks.
Why it matters
If you build AI agents or rely on frontier model reasoning, don't assume the model itself degraded—your provider may be serving a different inference regime. Instrument your API calls to log thinking tokens and latency, and compare against benchmark expectations before blaming prompt design.
Discussion angle
Should builders add telemetry to track thinking tokens and inference quality over time, and how do you handle production reliability when the same model can perform differently day to day?