Anthropic blocks users for anything touching training larger models
- ID
- 41674
- Status
- summarized
- Published
- 01 Nov 2026, 3:39 AM
- Fetched
- 01 Nov 2026, 2:18 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://twitter.com/jiahaoshao1/status/2116475922190922088
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 01 Nov 2026, 2:18 PM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
A researcher (Jiahao Shao, per the post's own bio a CS PhD student at HKUST) publicly accused Anthropic of flagging the account for "training a competitor," saying rollouts actually came from a strong open-source model and Claude was only used to submit jobs and run analysis. The claimed consequence: anything touching RFT, SFT, or training larger models now gets blocked, Opus 5.5 requests get silently downgraded to Opus 5, or the session is killed outright. The Hacker News thread drew 281 points and 152 comments, and a follow-up post argues the anti-distillation policy is a double standard given models trained on the open web.
Why it matters
This is an unverified user report, but the concrete claim to plan around is the silent downgrade: a request you think is going to Opus 5.5 can be served by Opus 5 or terminated, so any pipeline where output quality matters should log the model version actually returned and pin/assert it. If you use Claude for ML-adjacent automation — job submission, experiment orchestration, SFT/RFT data prep — assume provider-side classification of your workload can change your access without notice, and keep a second provider or a local open-weight model ready for those specific jobs rather than for everything.
Discussion angle
Where should the line sit between anti-distillation enforcement and blocking paid users doing legitimate research automation — and practically, how do you detect a silent model downgrade in production (logging returned model IDs, asserting on tier, alerting on fallback)?