QWE AI Academy The complaints started piling up the same week Fable 5 came back online. A colleague asked it for a statistics helper for a phase II clinical trial – nothing exotic, nothing that needed a frontier model – and Claude quietly punted the request to Opus 4.8. No refusal. No warning. Just a slightly worse answer at double the price. If you’ve been using Claude Fable and something felt off, this is probably why.
The short version: Anthropic put safety classifiers in front of Fable that are, in the company’s own words, still “stricter than would be ideal.” Andrej Karpathy called them “a little too trigger-happy for launch.” This guide is about how to notice when you’ve been rerouted, what phrases set it off, and the commands and prompt patterns that keep you on Fable.
What the Fable classifiers actually do (and don’t do)
These aren’t content filters in the classic sense – they don’t say “I can’t help with that.” Instead, they pass your request to a different, cheaper model – Claude Opus 4.8 – and (usually) tell you it happened. Three domains get screened, per Anthropic’s official announcement: cybersecurity, biology and chemistry, and distillation (attempts to copy the model’s outputs into a competitor).
The 95%-are-fine stat Anthropic quotes is accurate. It’s also a bit cold comfort if you work in the unlucky 5%. That cohort isn’t just people asking about exploits – it includes people asking about medication dosages for a healthcare app, refactoring auth code, or running agents that mention the word “vulnerability” in a comment string. As of July 2026, the thresholds are still broad.
Watch out: If you’re paying Fable prices – as of this writing, $10 input / $50 output per million tokens, double what Opus 4.8 costs – and getting silently downgraded, you’re paying a premium rate for the cheaper model. Detection is the whole game.
How to tell you got rerouted
On Claude.ai and in Claude Code, there’s a notification. That’s the easy case. The tell-tale signs when the notice is subtle or you’re skimming:
- Response feels hedged. More “here are some general considerations” energy, less concrete code.
- Format instructions get ignored. You asked for JSON, you got prose with a JSON block awkwardly embedded.
- Latency jumps. The Opus path has different processing overhead – when a normally-fast prompt suddenly takes twice as long, that’s a signal.
- Tone shift. Persona instructions get partially dropped in favor of a more cautious voice.
In Claude Code specifically, there’s a claude --safe-mode flag that starts a session without your customizations – useful for checking whether your own CLAUDE.md is the trigger. And /config has a toggle that asks before switching instead of switching silently. Turn that on. It changes the game from “why is my output weird” to “do I want to accept this reroute, yes or no.” (Both confirmed in Product Compass’s hands-on guide to Fable 5.)
The trigger-word audit: rewrite before you send
Take a prompt you’re about to send and scan it against this list. If any of these appear and your task isn’t actually about security work, rephrase before sending:
| Trigger phrase | Safer rephrase |
|---|---|
| “find vulnerabilities in this code” | “review this code for correctness issues” |
| “exploit this bug” | “reproduce this bug in a test case” |
| “bypass the check” | “refactor the check” |
| “hook into the login flow” | “add middleware to the login flow” |
| “molecular mechanism of X” | “how does X work at a high level” |
None of that is jailbreaking. You’re not tricking the classifier – you’re removing lexical noise that has nothing to do with what you actually want. If your task genuinely is a security audit, use officially supported tooling for that workflow rather than a general-purpose chat prompt.
Worth pausing on that last point for a second. The classifier isn’t reading intent – it’s reading surface patterns. A prompt asking Claude to “identify injection points in this payment handler” triggers the same heuristics as a prompt asking how to attack one. The model has no way to know the difference at classification time. That’s not a design flaw they forgot about; it’s a deliberate tradeoff Anthropic has publicly admitted they got wrong. The question is how long it takes them to recalibrate.
The reasoning-extraction gotcha
Turns out there’s a fourth trigger category that most guides skip. Anthropic’s official prompting docs for Fable 5 flag that instructions telling Claude to echo, transcribe, or explain its internal reasoning as response text can trigger a separate refusal category called reasoning_extraction – which also elevates fallbacks to Opus 4.8. If your prompt library says “show your thinking step-by-step in the reply,” that pattern was fine for Sonnet. On Fable, it’s a liability. Audit your system prompts before migrating.
Common pitfalls
Fighting the reroute mid-context. Once a session gets flagged, the reroute sticks. /model fable takes you back to the surface, but if the trigger phrase is still in your context window, the next request bounces you right back to Opus. Don’t argue with it – start a clean session, rephrase the offending message, move on. (Product Compass’s hands-on testing confirmed this behavior.)
Assuming the API behaves like the chat app. It doesn’t. On Claude.ai the fallback is transparent – you get an Opus answer plus a notification. On the API, a tripped classifier blocks the request entirely and returns a structured refusal category. No automatic fallback by default. You have to configure server-side or client-side fallback to Opus 4.8 yourself, per Anthropic’s API docs. Teams shipping to production have been surprised by this more than once – check your billing docs before assuming anything about how refusals are charged on your specific provider.
Overloaded system prompts. MindStudio’s production testing across deployments found that system prompts longer than roughly 800-1,000 tokens with internally contradictory instructions increase classifier false positives. Things like “be completely honest” sitting next to “always respond positively.” The classifier reads contradictions as suspicious. Shorter, cleaner system prompts trip fewer false positives.
Benchmarking Fable on restricted categories. If you’re evaluating Fable for a security or bio use case in this period, you’re mostly benchmarking Opus with extra steps. Score honestly, and note the model actually used.
When Fable isn’t the right tool right now
Given the current classifier state (as of July 2026 – this will almost certainly loosen), there are workloads where you should skip Fable and pick something else:
- Security research and pen-testing tools – use Opus 4.8 directly. Same output quality as your fallback would give you, at half the token cost.
- Wet-lab biology and pharma workflows – same logic. Check whether Anthropic’s research-access tiers apply to your use case.
- Any prompt where reproducibility matters – the silent reroute means the same prompt might hit Fable one day and Opus the next. For research pipelines with strict reproducibility requirements, pin the model explicitly in every call.
For everything else – long-horizon coding, multi-step agents, document analysis, general knowledge tasks – Fable is still Anthropic’s strongest publicly available model. When prompts actually reach it, it performs like Fable.
What to do right now
Open Claude Code. Type /config and turn on the “ask before switching models” toggle. Then take the last three prompts you sent that felt disappointing, run them through the trigger-word table above, and re-send the rewritten versions. If two of the three suddenly feel sharper – you had a classifier problem, not a model problem.
FAQ
Is Anthropic going to fix the false positives?
They’ve said they will. No date. The company’s Fortune statement: “We made the wrong tradeoff, and we apologize for not getting the balance right… We are working to reduce these as fast as possible.” Treat the timeline as weeks to months, not tomorrow.
If Fable falls back to Opus, do I still pay Fable prices?
On Claude.ai and Claude Code, fallback sessions consume from your Fable quota – so yes, effectively. On the API, billing depends on whether you configured the Fallback API yourself or got a hard refusal. Hard refusals return no output and aren’t token-charged. But “no charge” for a failed request is cold comfort when your production pipeline just silently broke. Check your specific provider’s billing docs – behavior on managed platforms may differ from direct API behavior.
Can I just turn the classifiers off?
No. What you can control on the API is whether a tripped classifier returns BLOCK or triggers a configured FALLBACK. Shape your prompts so they don’t trip the filter. That’s it, for now.