Anthropic ships two models with one set of weights — and routes your 'unsafe' prompt to a weaker one
Claude Fable 5 and Claude Mythos 5 share the same weights. The difference is a set of classifiers that, in flagged sessions, quietly hand your request to the older Opus 4.8 instead of the model you paid for.
Anthropic put one model behind two doors. Mythos 5 is the full-capability version for ultra-vetted clients; Fable 5 is the public one, wired to safety classifiers that silently route flagged queries to the weaker Opus 4.8 rather than refusing outright.
The benchmark jump is real on paper, but the loudest voice in the thread runs a bug-bounty program and isn't celebrating: he's watching for the day low-effort fuzzing turns into a real security edge.
And the 'free until June 22' rollout reads to many as a hook before usage-based billing arrives.
One model, two names, and a switch you can't see
Anthropic's launch post announces two models that share the same underlying weights. Claude Mythos 5 is the full-capability version, reserved for ultra-authorized clients. Claude Fable 5 is the one the public gets — the same brain, wrapped in safety classifiers. When those classifiers fire (cybersecurity, bio/chem, model-distillation, and competing-LLM development), Fable doesn't refuse: it silently routes the request to the next-most-capable model, Opus 4.8. Anthropic says this happens in fewer than 5% of sessions.
The capability claims are large. Fable is pitched as state-of-the-art on nearly every benchmark Anthropic tested, priced at $10 per million input tokens and $50 per million output — described as less than half the price of the Mythos preview. The post cites a Stripe codebase migration done in a day and the model playing Pokémon FireRed from vision alone. Subscribers on Pro, Max, Team and Enterprise get it free from June 9 to 22, after which access moves to usage credits.
The benchmark numbers, and the asterisk on one of them
The headline scores, reproduced from the announcement and collected in the thread:
| Benchmark | Mythos 5 | Fable 5 |
|---|---|---|
| SWE-bench Verified | 95.5% | 95.0% |
| Terminal-bench 2.1 | 88.0% | 84.3% |
| GPQA Diamond | 94.1% | — |
| Humanity's Last Exam (no tools) | 59.0% | — |
The number people fixated on came from Cognition's FrontierCode benchmark: Opus 4.7 at 5.2%, Opus 4.8 at 13.4%, Fable 5 at 29.3%. A huge jump — until a reader walked back the methodology.
"That blog post really makes it look like it's graded from an LLM's estimation of an OSS maintainer's review … a coding model can produce code that is just convincing enough to pass review even though it's actually totally wrong." — a commenter dissecting the benchmark's grading method
That matters because an LLM-estimated review score is exactly the kind of signal you can optimize against in reinforcement learning. The 29.3% may measure persuasiveness to a grader more than correctness.
"I am not looking forward to how it will empower the people who submit bug bounties to us"
The most-upvoted reaction wasn't about price or benchmarks. It came from someone who runs an offensive-security pipeline for a living.
"if this is what it claims to be, I am not looking forward to how it will empower the people who submit bug bounties to us … they still don't have a true security edge … I wonder if that's about to deeply change." — sebmellen, who runs a bug-bounty program
His point is precise: today's low-effort fuzzing produces duplicative, minor reports. A model that closes the gap between "ran a fuzzer" and "found a real bug" changes the economics of every security team that triages those reports.
The "free for subscribers" line read as a hook, not a gift
Several founders saw the rollout schedule as a funnel.
"The 'offer, then remove' aspect is a bit eyebrow-raising — it feels like they are trying to get subscribers to switch to usage-based billing, which makes me wonder if we'll ever get it after that June 22nd window." — a startup cofounder reading the billing change
A sharper version questioned the "ethical AI" branding directly:
"nothing says 'ethical AI' like only allowing mega corporations to use it through cost barriers." — Kyle Drake, founder of Neocities
The race nobody thinks they can leave
One voice put the release in the context of the whole field's momentum, and why no lab feels free to slow down:
"It's the 'If we don't, someone else will' effect … Half the comments on HN lately are 'wtf Claude is so dumb compared to Codex; I'm switching' — nobody can slow down while those exist." — a commenter on the competitive treadmill
The same crowd that demands more capability is the crowd whose threats to switch make restraint commercially impossible. Anthropic's own cited METR evaluation hedges the ceiling: Mythos 5 is "likely unable to fully and reliably automate R&D for frontier projects spanning multiple weeks." Two doors, one set of weights, and a model that decides — without telling you — which one you're talking to.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
This is the last story in the June 10, 2026 issue.
Go to the issue archive