Two editions ago, the story was that Claude Fable 5 was finally coming back after a month locked up by US export controls. Now that it's here, a lot of people are saying the thing that returned isn't the thing that left. The word going around is "nerfed." The real explanation is more interesting than that.
What people are seeing
Within a day of Fable 5's return, developers started reporting that it felt gutted. "The new guardrails are kicking in on way too many tasks," one wrote, adding that "this is not the model that got banned." A systems programmer found the model balking at ordinary work the moment it touched C, C++, Rust, or files that so much as mentioned "security," "vulnerable," or "hook."
Then the benchmarks landed and looked ugly. A community testing group, BridgeMind, ran the relaunched model through its coding suite and watched debugging and refactoring scores fall off a cliff compared with the June version.
What's really going on
As a condition of getting Fable 5 back on the market, Anthropic added an aggressive safety classifier aimed at stopping the exact jailbreak that triggered the government's ban. When that classifier decides your prompt looks even a little like offensive security work, it doesn't refuse. It hands your request to the older Opus 4.8 instead.
That is why the benchmarks cratered. In one run of twelve coding tasks, only three actually reached Fable 5. The other nine were intercepted and scored as zeros, so the "Fable 5 score" was mostly a measure of the backup model getting questions that were never meant for it.
The calmer data
Not every test agreed. Arena.AI collected thousands of blind human-preference votes and found the picture far less dramatic: coding slipped a little, but writing, documents, and expert text held steady or came out ahead of the June version. The takeaway is that this is task-specific. If you write, research, and analyze, you are mostly fine. If you lean on Claude to fix and debug code, you are the one hitting the tripwire.
Silent or announced?
The sharpest disagreement is over whether the model tells you when it swaps. Many users say the handoff to Opus 4.8 gave them no warning, which is what made it feel like a downgrade. Anthropic's Sholto Douglas pushed back directly, saying the model only ever falls back "loudly," and that an earlier silent version was reverted. Both things can't be fully true at once, and Anthropic hasn't shown the receipts either way, so treat the "secret downgrade" version as a claim, not a fact.
What both sides agree on: the company has acknowledged the classifier "casts too wide a net" and says it will refine it, without giving a timeline.
We spent two editions on Fable 5 coming back, and the real worry going in was that it would return as a watered-down version of itself. It didn't. The model itself looks just as strong as the original launch, and almost nobody is arguing otherwise.
What's left is the safety filter, and the panic there is overblown for most people. In my own use I haven't been bumped to Opus 4.8 once.
If you write, research, or make things, you're almost certainly fine. If you work near code, the one thing worth keeping an eye on is whether a security-sounding word is rerouting you to the older model.
Anthropic already calls the filter too aggressive and says it's dialing it back, so even that should ease up.
Sources
- Decrypt: Claude Fable 5 Isn't Nerfed. The Router Is Just Paranoid - Jose Antonio Lanz - BleepingComputer: Claude Fable relaunch disappoints users with nerfed performance - Sholto Douglas (Anthropic) on the fallback being loud - Theo on worse iOS coding - Modem Guides: Is Claude Fable 5 Nerfed? What the Benchmarks Really Show