Claude Fable 5 - is this Mythos model worth the wait?

Overview

• Anthropic launched “Mythos” as a new model class, specifically “Claude Fable 5” (often framed in the video as “baby Mythos”).
• The creator received early access and tests it with a focus on:
  • whether it matches marketing hype
  • whether it’s practical for their own “backlog”
• The model is described as state-of-the-art on benchmarks and strong on long, complex, autonomous tasks, but the reviewer flags cost, token intensity, and quality issues in some real-world workflows—especially specs and design.

Product positioning & safety / restrictions

• “First Mythos-class intelligence model to reach GA”, but the video emphasizes that Fable 5 isn’t full “Mythos (capital M)”:
  • It includes guardrails for cybersecurity and biology (and possibly other sensitive areas).
• Safety classifiers target categories like:
  • cyber security
  • biology
  • chemistry
  • (and “distillation” is mentioned)
• Fallback behavior (key feature):
  • If classified into restricted categories, it falls back to Opus 48 instead of hard-blocking.
  • The API supports graceful fallback to Opus 4.8 / Opus pricing for continued handling.
• Retention / misuse handling:
  • A 30-day retention policy is used for misuse checking, explicitly stated as not used to train Claude.
  • The reviewer notes that most sessions don’t trigger fallback (claims ~95% did not).

Benchmarks & performance claims

• Anthropic claims Fable 5 exceeds benchmarks by a significant margin.
• A cited metric: ~80% on SweetBench Pro (compared to other recent models).
• Compared against: Opus 4.8, GPT-5 (“GPT55” as written), Gemini 3.1 Pro—the video claims Fable 5 is far ahead on the tested benchmark(s).
• Strength areas Anthropic emphasizes:
  • long, complex tasks
  • autonomy, including asynchronous / days-long task style
  • agentic execution, including sub-agents / dynamic workflows
  • vision
  • “effort”/thoroughness (verifies more)

Cost / efficiency notes (from the reviewer)

• Pricing is described as very expensive:
  • $10 per input token
  • $50 per output token
• The reviewer claims token usage is ~2x compared to other models (both faster and higher consumption are implied).
• They explicitly recommend paying attention to:
  • cost
  • efficiency
• Testing was often done at higher “effort levels,” which increases token burn.

What it’s good at (reviewer’s hands-on findings)

1. Vision + document formatting

   • The reviewer is genuinely impressed.
   • Strong at:
     • PDF/document parsing
     • layout and formatting
   • Example mentioned: formatting a handwriting sheet for a child’s work—spacing and readability improved versus other models (an Opus vs Fable 5 comparison is referenced).

2. Long-horizon / agentic workflows (partially validated)

   • The reviewer could get hours-long runs.
   • They did not personally run it “for many days”, so they couldn’t fully verify the “days” promise, but it seemed capable.

3. Hard technical problems (where detail matters)

   • The reviewer suggests using it for:
     • detailed engineering / technical reasoning
     • ambitious execution
   • They also note that multi-agent orchestration produced some successes.

What doesn’t work as well (reviewer’s critiques)

1. “Too engineer-like” thoroughness hurts spec/PRD readability

   • Output is:
     • complete and detailed
     • but hard to “zoom out” and parse
   • Specs/reviews can look “long and intelligent” yet become nearly impossible to read as high-level product documents.

2. Bad one-shot UI/design quality

   • When asked to design a skills registry UI, the result was fundamentally terrible design (not just generic “AI slop”).
   • Even with more detailed prompting, the design remained disappointing.
   • Conclusion: don’t rely on Fable 5 for front-end design.

3. May be too conservative about execution/MVP ambition

   • When asked to ship an MVP that gives value to a customer, it produced something too minimal / narrow.
   • This is framed as potentially related to safeguards or general tuning toward caution.

4. Multi-agent orchestration can stall

   • The reviewer experienced stalls/errors after leaving dev agents running for ~3 hours.
   • They suspect some issues might involve Claude Code / tooling rather than the model, but emphasize that long-running agentic systems must work reliably.

Recommended usage pattern (implied “best fit”)

• Use Fable 5 for:
  • vision + document formatting
  • hard technical tasks
  • long-horizon, agentic execution
  • cases where thoroughness/detail is needed
• Prefer other models (e.g., Sonnet / Opus) for:
  • spec writing / PRDs / strategy docs (it over-investigates and outputs dense text)
  • front-end design
• Potentially use it as:
  • an “advisor” while cheaper models execute (described below under product features).

Launches / API & product features mentioned

• CloudManage Agents: described as a hosted harness/sandbox for running long-running agentic work.
  • The reviewer is unsure of the best use cases but notes Fable ships with them.
• New “advisor strategy”:
  • Use Fable 5 as a senior adviser.
  • Use cheaper models as an execution layer.
  • Works in both API and “cloud code.”
• Fallback API parameter:
  • Allows optional fallback to Opus pricing when requests get blocked/safelisted.

Final takeaway (reviewer’s conclusion)

• The model is not dismissed (“not a hater”).
• Best described as:
  • valuable for hard problems and vision/document formatting
  • less ideal for specs/strategy and design
  • potentially expensive and token-intensive
  • useful but requires careful model routing (which model + effort level + task type)

Main speaker/source

• Creator/host: “How I AI” (the narrator who runs the tests and gives personal evaluation)
• Source referenced: Anthropic (model release claims, benchmark and safety/fallback details)