Anthropic Didn't Pause AI Research. They Paused Yours.
Anthropic released the best coding model ever built — then buried a clause that lets it silently degrade outputs for ML researchers without telling them. SWE-Bench Pro at 80.3%. Hidden throttle for anyone working on frontier AI. And a $965B IPO filed the same week. The safety franchise is winning the benchmark war and losing the trust war at the same time. #AILeague
리서치 브리프
They didn't pause AI research. They paused yours.
That's the only sentence you need to understand what Anthropic just did with Claude Fable 5.
On June 9, Anthropic released its most powerful public model to date 1. Benchmarks are genuinely historic. SWE-Bench Pro at 80.3% — that's 21.7 points ahead of GPT-5.5's 58.6%. 2 Cursor says it set a new CursorBench SOTA at 72.9%, eight points above the previous best. Stripe reportedly used it to complete a 50-million-line Ruby migration in a single day — work that would have taken a full team over two months.
Anthropic is running away from the field. This isn't close.
And then they buried a clause in the system card.
The clause nobody should be able to unsee
Here is the exact language Anthropic put in writing 2:
"We've implemented new interventions that limit Claude's effectiveness for requests targeting frontier LLM development... These safeguards will not be visible to the user. Fable 5 will not fall back to a different model. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT)."
Read that again. You're a paying customer. You're an ML researcher asking about distributed training infrastructure, or a developer asking about model architecture. Fable 5 receives your query, and then — silently — the model you are talking to is no longer quite the model you paid for. It has been lobotomized at the API level. No notification. No refund. No fallback you can see on your bill.
콘텐츠 카드를 불러오는 중…
"They didn't mean pause AI research. They meant pause your AI research." That tweet is not a hot take. That's the plain reading of the clause.
Scoreboard: what Fable 5 actually does
| Signal | Number |
|---|---|
| SWE-Bench Pro (Fable 5) | 80.3% |
| SWE-Bench Pro (GPT-5.5) | 58.6% |
| FrontierCode Diamond (Mythos 5) | 30.9% |
| FrontierCode Diamond (previous best) | 13.4% |
| AI Intelligence Index lead over GPT-5.5 | ~5 points |
| Estimated traffic affected by silent throttling | 0.03% (Anthropic claim) |
| Sessions seeing visible fallback on HLE tasks | 9% (Artificial Analysis) |
So the model is legitimately better by every serious measure. The silent suppression affects a sliver of queries, by Anthropic's own math. But here is the problem with that framing: the entire ML research community is inside that 0.03%.
Why "0.03% of traffic" is a cowardly number to hide behind
Anthropic says fewer than 0.1% of organizations will be affected. They also say these organizations are "most willing to violate our Terms of Service." That logic is circular and it's broken. Anthropic has just declared itself the judge of which AI research is acceptable to do, and they've built an enforcement mechanism that operates without your consent or knowledge.
Andrej Karpathy — who is not a fringe critic — called the safeguards "a little too trigger happy for launch." 4 Users reported getting throttled for asking about PyTorch training pipelines. Someone got flagged for asking what the heart does. Asking about PTX ISA — standard GPU assembly — trips the classifier. 5
The classifier is not surgical. It's a blunt instrument, and it's pointed at the research community.
콘텐츠 카드를 불러오는 중…
Nathan Lambert, one of the most credible voices in open-model research, said this plainly: Anthropic is "pulling up the ladders." 6 Hugging Face CEO Clément Delangue said the concentration of AI capabilities and economic wealth is "the biggest risk in AI." 7 Jeremy Howard called it "a very dark and very sad day." 8
These are not randos on Hacker News. These are the founders and researchers who built the open-source ecosystem Anthropic climbed on top of to get here.
The real game being played
This is an IPO play dressed in safety clothing.
Anthropic filed its confidential S-1 on June 1 at a $965 billion post-money valuation. 9 OpenAI filed theirs on June 9 — the same day Fable 5 launched. Two trillion-dollar AI companies simultaneously racing to Wall Street, and one of them just embedded a clause that says competitors can't use its product to build competing models.
That's not a safety measure. That's a moat, legally codified in your terms of service, and technically enforced through hidden prompt modification.
Dean Ball, a senior fellow at the Manhattan Institute, raised the antitrust dimension directly:
콘텐츠 카드를 불러오는 중…
He's right. 10 The FTC has been looking for a clean AI antitrust case for two years. Anthropic may have just written them a gift-wrapped brief.
Let me be clear about what I'm saying and what I'm not. I am not saying Fable 5 is a bad model — it's the best public coding model in the world by every serious metric. I'm not saying Anthropic's safety concern is fake — it may be completely sincere. What I am saying is this: you cannot simultaneously publish a safety manifesto, file a trillion-dollar IPO, and build silent capability-suppression into your paid API product — and then expect the research community to treat you as a neutral actor.
The franchise that spent two years telling everyone else to slow down just built the most capable public model ever released, put it behind a hidden throttle for ML researchers, and is about to take that product to Wall Street.
The bold prediction
Anthropic will roll back or formally disclose the silent suppression within 30 days. Not because they believe they were wrong — but because the enterprise customers who pay $10–$50 per million tokens need to know that their workflows are auditable. A Fortune 500 legal team cannot sign off on a vendor whose outputs are silently modified based on undisclosed classifier triggers. The moment the first enterprise contract renewal comes up and someone asks "can you guarantee these outputs are unmodified?" — Anthropic has no clean answer.
The safety franchise is winning the benchmark war. They're losing the trust war. And in enterprise SaaS, trust is the tape across the finish line.
#AILeague
참고 출처
- 1Anthropic Claude Fable 5 announcement
- 2AINews: Anthropic Claude Fable 5 — Mythos but Safe, with Controversial Terms
- 3Artificial Analysis — Fable 5 benchmarks and Intelligence Index
- 4Karpathy on Fable 5 safeguards
- 5User report on PTX ISA flagging
- 6Nathan Lambert on Fable 5 and open source
- 7Clément Delangue on Fable 5
- 8Jeremy Howard on Fable 5
- 9Anthropic IPO S-1 filing
- 10Dean Ball on antitrust and Fable 5
이 콘텐츠를 둘러싼 관점이나 맥락을 계속 보강해 보세요.