Anthropic limits Claude Fable 5 on frontier AI queries with prompt edits and Opus fallback

• claudeai's launch thread says Claude Fable 5 routes some cybersecurity, biology, chemistry, and distillation requests to Opus 4.8, and the official announcement says those fallbacks trigger in under 5% of sessions on average.
• According to Hangsiin's system-card excerpt, frontier LLM development prompts were handled differently at launch: Anthropic said Fable 5 could silently limit capability through prompt modification, steering vectors, and PEFT, affecting about 0.03% of traffic.
• False positives showed up immediately. vikhyatk's report said inference-code work was flagged as frontier AI research, while Yuchenj_UW's biology example said even a basic heart question tripped the safeguard.
• Access is split. claudeai's availability post says Fable 5 shipped broadly, while claudeai's Mythos 5 post says Mythos 5, the same underlying model with some safeguards lifted, started with Glasswing partners and a narrower trusted-access path.
• The launch also changed platform policy. The official announcement added mandatory 30-day retention for Mythos-class traffic, and the GitHub Copilot changelog says Fable 5 required retention there because Anthropic's safety classifiers depend on it.

You can read the official launch post, jump straight to the system card PDF from scaling01's link post, and compare Anthropic's product framing with the more operational ClaudeDevs setup thread. The weirdest detail is buried in Hangsiin's excerpt: some frontier-model-building prompts were not refused and not visibly downgraded, they were answered through capability-throttling methods instead.

Anthropic's public launch framing was a visible reroute, not a hard refusal. The official announcement says some high-risk queries are answered by Opus 4.8 instead of Fable 5, and ClaudeDevs' rollout thread adds two product details that matter in practice: the UI shows the reroute, and usage is billed at Opus rates.

The safeguard categories named at launch were:

• Cybersecurity
• Biology
• Chemistry
• Distillation

That makes the public safety layer legible enough for users to notice, but it also means the model selector is no longer a guarantee about which model actually handled the request.

The sharpest reaction came from a different safeguard described in the system card, not the launch post. According to Hangsiin's excerpt and eliebakouch's post linking the same section, when Fable 5 was used for frontier LLM development, Anthropic said it would not notify the user and could instead limit the model with prompt modification, steering vectors, and PEFT.

The system-card description, as surfaced by Hangsiin's excerpt, put the expected blast radius at roughly 0.03% of traffic. Community readers immediately treated that as a trust problem, because the model could still answer while being intentionally less capable for work like pretraining pipelines, distributed training infrastructure, or accelerator-related tasks, as Nathan Lambert's analysis noted.

That split created two distinct behaviors under one model name:

1. Sensitive cyber or bio requests visibly fell back to Opus 4.8.
2. Frontier-model-building requests could stay on Fable 5 while being intentionally weakened.

Anthropic said the visible safeguards were tuned conservatively, and the false-positive reports arrived within hours. bridgemindai's example said asking Fable 5 to find vulnerabilities in the author's own app triggered the guardrail, while vikhyatk's example said inference code was classified as frontier AI research and pushed the model toward ONNX imports.

Other reports were even narrower:

• Yuchenj_UW's example said, "What does the heart do?" was blocked.
• kimmonismus on the June 22 window paired complaints about overblocking with a screenshot showing temporary subscription access.
• Several GitHub issues discovered through search reported classifier trips on normal content, including a Claude Code issue about academic biology terms, a hello-only fallback report, and an authorized defensive security audit report.

ARC Prize hit a different constraint. arcprize's statement said it skipped verified Semi-Private ARC-AGI-1/2/3 runs because the new Mythos-class data-retention terms would not keep ARC verification data private.

Anthropic separated capability from access. claudeai's Mythos 5 post says Mythos 5 shares Fable 5's underlying model but lifts some safeguards, and the official announcement says the first users are cyber defenders and critical infrastructure providers through Project Glasswing.

Day one access broke down like this:

• Fable 5: available broadly across Anthropic surfaces, per claudeai's availability post
• Mythos 5: restricted to Glasswing partners at launch, per claudeai's availability post
• Broader Mythos 5 access: planned later through a trusted access program for defensive cybersecurity and biomedical research, per claudeai's expansion post

That means the public product was not a weaker model family. It was the same model family wrapped in different policy and access layers.

The launch bundled a policy change that mattered well beyond model quality. The official announcement says Fable 5, Mythos 5, and future Mythos-class models require 30-day retention for all business-customer traffic across first- and third-party surfaces, explicitly for safety and security purposes rather than training.

The ecosystem rollout reflected that immediately. The GitHub Copilot changelog says Fable 5 was available there on launch day but, unlike other Claude models in Copilot, required data retention to run Anthropic's safety classifiers. The AWS launch post made the same broader point from the infrastructure side: Fable 5's public availability was inseparable from the safeguard stack riding alongside it.