Claude Fable 5.1 and Claude Mythos 5.1
Anthropic’s release of its high-end Claude Fable 5.1 and Mythos 5.1 models prompts mixed reactions: some users report clear gains in reasoning, long-horizon coding, and scientific work, while many question whether the marginal improvements justify the high cost and strict safeguards. Commenters highlight frustrations with aggressive safety filters that silently downgrade to weaker models, opaque and tightening usage limits, verbose and stylized prose, and new watermarking and “preserved thinking” constraints that feel like model DRM. Others note that rapidly improving, much cheaper open and Chinese models are eroding Anthropic’s pricing power, pushing many developers to fall back to older Claude versions, OpenAI’s Sol, or low-cost alternatives for everyday work.
Pricing, Quotas, and Usage Limits
- Many users see Fable 5.1’s cache-read discount as good in theory but note that real-world task cost can be flat or higher vs Fable 5, especially at high effort.
- Subscription users are frustrated that cache pricing changes seem API-only and that subscription usage accounting is opaque.
- Several complain that the 20x “Max” plan is effectively far less than 20x, that weekly limits are easy to hit (especially with Fable), and that recent “resets” actually lowered effective usage vs the temporary bump.
- Some say they’ve cancelled or downgraded Claude subscriptions and now mix Anthropic with OpenAI, GLM, DeepSeek, Kimi, etc., often citing better quota and cost per task.
Safeguards, Cyber/Bio Work, and Mythos
- A dominant theme: Fable is often unusable for security, kernel, networking, auth, or some bio/science workloads because it silently or explicitly downgrades to Opus or hard-refuses.
- Even users in Anthropic’s cyber program report that cyber-unlocked access doesn’t apply to Fable; only Mythos gets full capabilities, and that’s tightly gated.
- Others say safeguards have noticeably relaxed since early Fable 5 and now only occasionally trip, but this seems domain-dependent.
- Many argue these guardrails mostly hurt defenders and legitimate researchers while not truly stopping bad actors.
Writing Style and “Claudisms”
- Very heavy criticism of Opus 5 and Fable 5 prose: verbose, dense, full of distinctive turns of phrase, long comments, and “TED talk” tone. People call it exhausting for long coding sessions and PR reviews.
- Some say Fable 5.1 is indeed more readable and less mannered, but others report it’s still dense and that Anthropic is patching symptoms via prompts instead of fixing post-training.
- Users try output styles, CLAUDE.md, and hooks to enforce brevity; these often decay after a few turns.
Watermarking and “Preserved Thinking”
- The new text watermarking (per EU AI rules) draws concern: fear of false positives in plagiarism/AI-detection and discomfort with “secret codes” in their code/docs.
- Anthropic claims no quality loss; many remain unconvinced but lack concrete counterexamples.
- The “Preserved Thinking” / append-only transcript rules and anti-distillation changes are seen as “model DRM,” breaking existing patterns (editing system prompts, tool formats) and making custom harnesses harder.
Model Quality and Comparisons
- Opinions split: some say Fable 5/5.1 is a huge step up for complex coding, formal methods, math, and science; others say Opus 4.6/4.8 or GPT-5.6 Sol feel more reliable, less slop, and better value.
- Multiple people feel frontier gains are now marginal; the exciting frontier is cheap, “good-enough” models (GLM 5.3 Flash, DeepSeek V4 Flash, Luna, etc.) and local models.
Trust, Support, and Strategy
- There is visible distrust toward Anthropic: billing glitches, slow support, opaque safeguards, and a sense that employees use internal Mythos variants and don’t feel the pain of public models.
- Some see Anthropic as over-optimizing for safety, IPO optics, and IP protection at the expense of usability and openness.