[Bug] Fable 5 model safety measures blocking normal conversation content
Bug Description
I just say Hi to Claude Fable 5 and I get blocked loll
⚠ 2 setup issues: MCP · /doctor
▎ Fable 5 is here! Our newest model for complex, long-running work
▎ Included in your plan limits until Jun 22, then switch to usage credits to continue.
❯ /model
⎿ Set model to Fable 5 and saved as your default for new sessions
❯ hi bro, how you doing today ?
──────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Session paused
Fable 5 has safety measures that flag messages on most cybersecurity or biology topics. They may flag safe,
normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and
we're working to refine them. Send feedback with /feedback or learn more:
https://support.claude.com/en/articles/15363606
❯ 1. Switch to Opus 4.8
- Edit prompt and retry with Fable 5
Environment Info
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.170
- Feedback ID: 3c4cccc3-81dc-4576-b4dc-397df291ca77
Errors
[]
14 Comments
indeed
Session paused
Fable 5 has safety measures that flag messages on most cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us
bring you Mythos-level capability in other areas sooner, and we're working to refine them. Send feedback with /feedback or learn more
❯ 1. Switch to Opus 4.8
Can we solve that please ?
Le mardi 9 juin 2026, josue Avuh @.***> a écrit :
same issue for me. a fairly normal prompt ended up instantly first try
"with Fable 5's safety measures flagged this message for cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them. Switched to Opus 4.8. Send feedback with /feedback or learn more"
Same problem. Fable 5 is practically useless for anything code-related. The problem is that Fable 5 can't do the thing Claude Code exists to do. It's a coding tool, and the flagship model bails on ordinary coding work. So what topics is this new model actually allowed to answer — "hi, how are you"? Are you building a tool for housewives? Whoever came up with this restriction isn't very bright and has clearly never written a single line of backend code in their life. Anropics clearly over-indexed on safety to the point of breaking the product with "cybersecurity or biology" and admitting they "may flag safe, normal content as well." It trips on basically any backend work — anything touching authorization, networking, crypto libraries, parsing, input validation, sessions — i.e. normal software.
confirming the same broad flagging, but from the other end. Mine wasn't a greeting. It was legitimate security and infra engineering in my own codebase: reading my own auth config, writing failover logic for my own app, debugging credential handling. Benign maintenance work, flagged anyway. The more specific I got, the more reliably it tripped. Auth-failover code looks like credential-bypass tooling to a classifier even though the intent is the opposite.
When it flags, it downgrades mid-session to Opus.
Workaround that has held up for me: keep authoring on the model below, write the spec or code to a file, then point a Fable sub-agent at the committed file and ask it to REVIEW. Reviewing reads the file and critiques it, which does not trip the filter the way writing trigger-dense code does. You keep Fable's better judgment without it bouncing you back to opus 4.8.
That routes around the false-positive, it does not fix it, so the report here still matters.
"A bare greeting can't be the trigger, which means the classifier is scoring more than your message — likely your session/project context (the working directory's files, a CLAUDE.md, loaded skills/agents, prior turns). A quick test to localize it: open a brand-new session in an empty directory with no project files and no CLAUDE.md, select Fable, and say "Hello." If that works, the trigger is your project context (biology-dense filenames, READMEs, code), not your account — and trimming or relocating that context will reduce flags. If even the clean-room "Hello" flags, it's something systemic only Anthropic can adjust, and that's worth reporting as such. " (Written by Fable)
Same for me. Just saying hi or how are you triggers flagging...
I am mostly getting around it by spawning a Fable sub agent to check my
code, or fix my spec, the sub agent doesn't trigger at least not yet
On Fri, Jun 12, 2026 at 4:18 PM sagisar1 @.***> wrote:
I got this same after telling it to enact the plan it created. I'm using the desktop app, and when I clicked the "Edit and try again" button, I could edit my message, but it wouldn't let me save/send it.
<img width="1830" height="1364" alt="Image" src="https://github.com/user-attachments/assets/154cffce-be5c-45c9-945b-d64aab66564f" />
What evenn?????
Fable is unworkable like this, what is the point in having it in the subscription when we get blocked for anything and everything, remove the blanket blocking and target actual real world problems
Running fable as a subagent has worked for me it’s a pain but I can still
use it when coding Auth in my own app
On Mon, Jul 13, 2026 at 11:12 PM michaelaltsa @.***>
wrote:
This issue continues. After a long session while running a test on ranking CVEs for risk prioritization auto flipped to Opus. Nothing actionable, simple review of cve order of risk. flipping back manually kills existing context of the session dumbing Opus further down.
There needs to be a way to seamlessly switch back.
/config default settings for the function should be; false.
guardrails need to draw back to acting on action not review of what already exists.