CHAPTERS
- 0:00 – 1:30
Fable 5 arrives: “baby Mythos” and the real question—can it ship your backlog?
Claire introduces Anthropic’s new Claude Fable 5, framed as the first generally available “Mythos-class” model—though not the fully unrestricted “capital-M Mythos.” She sets expectations: it’s benchmark-crushing and hyped, but the review will focus on practical developer/product work and whether it actually helps ship.
- •Fable 5 positioned as the long-awaited Mythos-class release (with caveats)
- •Central evaluation criterion: real-world productivity vs marketing hype
- •Claire has early access and will share hands-on pros/cons
- •Immediate framing: everyday software/prd workflows, not sensational use cases
- 1:30 – 3:02
Anthropic’s claims: new model class, autonomy, vision, and “seasoned engineer” behavior
Claire summarizes what Anthropic says Fable 5 is: a new model class beyond Sonnet and Opus, designed for long, complex tasks and autonomy. The model is pitched as proactive, high-effort, and notably strong at vision—traits that can be helpful but can also create product friction.
- •Mythos becomes a new tier/class with Fable 5 as the first GA release
- •Claimed strengths: long-horizon complexity, autonomy, proactive behavior
- •“Engineer’s engineer” positioning: thorough verification and investigation
- •Vision performance highlighted as exceptional
- •Tradeoff hinted: thoroughness doesn’t always serve shipping/product clarity
- 3:02 – 3:33
Token intensity and cost: paying for “big boy” reasoning
The model is expensive and intentionally token-hungry, consuming rate limits/tokens faster than other models. Claire notes she ran many tests at extra-high effort to avoid underpowering the model, raising the broader question of whether the additional spend actually yields better outcomes.
- •Pricing called out as premium tier above Opus
- •Anthropic claims ~2× token/rate consumption vs other models
- •Extra-high effort burns tokens quickly; “high” may be the practical sweet spot
- •Core evaluation: does token intensity translate into better results?
- •Need for humans to match model/effort level to task complexity
- 3:33 – 6:34
Days-long tasks and agentic workflows: promise vs practicality
Anthropic positions Fable 5 as capable of days-long asynchronous work, sub-agents, and multi-day sessions. Claire can’t fully validate multi-day runs but does observe the harness and model appear capable of sustained multi-hour work—sometimes beyond what the task warranted.
- •Claimed capability: days-long planning and asynchronous execution
- •Support for sub-agents and dynamic workflow architectures
- •Claire observes multi-hour endurance, but questions appropriateness for some tasks
- •Emphasis on long-running sessions as a key differentiator
- 6:34 – 7:35
Safety classifiers + graceful fallback: how Fable stays “safe enough”
Claire explains Anthropic’s added safeguards: classifiers for cyber, bio, chemistry, and distillation-related misuse. Instead of hard refusal, Fable can “fallback” to Opus 4.8, and Anthropic adds a limited retention policy intended to detect misuse without training on the data.
- •Safety classifiers target high-risk domains (cyber/bio/chem/distillation)
- •New concept: graceful fallback to Opus 4.8 instead of full blocking
- •Fallback is available as an API capability/parameter
- •30-day retention for misuse detection; not used for training
- •Anthropic reports most sessions do not trigger fallback
- 7:35 – 8:36
Is Fable 5 actually Mythos? Understanding “Fable vs Mythos” access tiers
Claire clarifies the naming confusion: Fable is Mythos-class but includes safeguards and is generally available; “Mythos” proper remains restricted to select Project Glasswing partners. She frames them as the same underlying model, segmented by access and guardrails.
- •Fable = Mythos-class with safeguards; Mythos (unrestricted) remains gated
- •Mythos access limited to Project Glasswing/enterprise partners
- •Practical implication: most users get “baby Mythos” today
- •Expectation of possible future expansion (versions or broader access)
- 8:36 – 9:06
New launches alongside the model: Managed Agents, advisor strategy, and fallback API
Anthropic ships product updates with Fable 5, including Claude Managed Agents (hosted agent sandbox) and a recommended advisor/executor pattern that pairs Fable with cheaper models. Claire also notes the new fallback API mechanism that continues execution on Opus 4.8 when needed.
- •Claude Managed Agents enters public beta; Fable available out of the box
- •Advisor strategy: Fable as senior advisor with cheaper execution models
- •Fallback API parameter enables graceful downgrade behavior in production
- •Claire is still exploring compelling Managed Agents use cases
- 9:06 – 10:06
Benchmarks: big SWBench Pro jump and “state-of-the-art” positioning
Claire reviews the benchmark claims and charts, especially the strong SWBench Pro score relative to other top models. While her tests weren’t the most extreme, she didn’t find clear technical failures, making her more inclined to believe the benchmark story.
- •Benchmark focus: SWBench Pro and broad across-the-board gains
- •Claimed lead over Opus 4.8 and other frontier competitors
- •Claire’s anecdotal validation: didn’t hit obvious hard failures
- •Positions Fable 5 as Anthropic’s new SOTA offering
- 10:06 – 11:37
Hands-on strengths: vision and document formatting that actually looks better
In real use, Claire’s standout positive is vision—specifically, producing cleaner, more readable document layouts. She compares outputs and finds Fable 5 better at spacing and formatting for a practical PDF/worksheet-style artifact.
- •Best-in-review capability: vision applied to document/PDF formatting
- •Improved layout: spacing, whitespace, readability, visual clarity
- •Practical example: handwriting worksheet formatting vs Opus output
- •Claire notes this is a simple eval but consistently impressive
- 11:37 – 13:09
Test 1 (Product graph spec): thoroughness that becomes unreadable prose
Claire tests Fable 5 on an adversarial review of a complex product graph spec. The output is detailed and extensive but difficult to parse—dense blocks, heavy internal references, and poor zoomed-out clarity—making it less useful for PRDs/spec communication.
- •Used for adversarial requirement review on a complex open-source project
- •Output: long, “intelligent-looking” markdown that’s hard to read
- •Key failure mode: can’t see the forest for the trees
- •Suggestion: use simpler models for specs; use Fable where you don’t need to read everything
- 13:09 – 14:10
Test 2 (Skills registry UI): surprisingly poor design and heavy prompt-dependence
When asked to design a skills registry, Claire finds the one-shot UI design output unusually bad—color choices, layout, and overall quality. Even with more detailed prompts (per Anthropic’s suggestion), results remain unimpressive, leading her to recommend other models for front-end/design tasks.
- •One-shot UI design quality described as fundamentally terrible
- •Prompting needed more specificity than expected for modern models
- •Even with improved prompting, design results still lagged
- •Recommendation: avoid Fable for front-end/design; consider Opus/others
- 14:10 – 15:41
Conservative execution and multi-agent orchestration: MVP too minimal + tooling stalls
Claire finds Fable 5 can be overly conservative when executing toward an MVP, producing outputs that are narrowly “minimal” rather than valuable. She also tests multi-agent/dynamic workflows: capability is there, but stalls and errors (likely in Claude Code/tooling) undermine the promise of long-running orchestration.
- •Execution mode skews overly conservative; MVP interpretation too narrow
- •Possible influence from safety tuning/guardrails on ambition
- •Multi-agent runs can succeed but reliability issues appear
- •Observed stalls after hours when stepping away; questions technical readiness
- •Distinguishes model capability vs Claude Code orchestration/tooling issues
- 15:41 – 17:24
Final takeaways: where Fable 5 belongs in your stack (and where it doesn’t)
Claire concludes Fable 5 is best reserved for hard technical problems, long-horizon detailed work, and vision/document tasks. She advises against using it for front-end design and for strategy/spec writing due to overthinking and impenetrable prose, and suggests consulting Anthropic’s prompting guide for better outcomes.
- •Use for hardest technical problems where detail and verification matter
- •Use for vision-heavy tasks: parsing docs, formatting, visual outputs
- •Avoid for front-end/design and for strategy/spec prose (overly dense)
- •Consider lowering effort level for prose experiments; mix models intentionally
- •Closing: Mythos-class is here; encourages viewers to experiment and share results
