Video summary

MYTHOS MYTHOS MYTHOS

Main summary

Key takeaways

Technology

Anthropic’s Mythos 5 & Fable 5: “a new class of AI”

Anthropic’s Mythos 5 and Fable 5 are presented as a major step up from prior Claude “generally available” models. The speaker frames them as a new class of AI with markedly different behavior.

The speaker claims early access to Mythos 5 and argues that Fable/Mythos are “frontier” in capability and autonomy—especially on long-horizon, complex tasks.

Model lineup & key distinction

Claude Mythos 5 vs Claude Fable 5

  • Only difference: Fable has guard rails; Mythos has guard rails removed.
  • Mythos (guard rails removed) is positioned for the security community, e.g.:
    • hardening software
    • finding bugs
  • Fable (guard rails on) is positioned for general use.

The model is described as extremely capable, “feels like a new training run.”

The speaker also claims that a Mythos variant described as “too dangerous to release publicly” was ultimately released publicly anyway.

Benchmarks (and critique of them)

The video focuses heavily on reported benchmarks, while warning that benchmarks can be misleading and may not match the “vibes” experienced in real use.

Reported evaluation highlights mentioned include:

Coding (Aider SWE-bench / similar coding benchmarks)

A reported Aider SWE-bench Pro comparison:

  • Mythos 5 (or related frontier grouping): ~80%
  • Claude Opus 4.8: ~69%
  • GPT-5.5: ~58%

The speaker notes that benchmark ordering can feel inconsistent with real perceived capability. For example, GBT/“GPT” is said to be more capable than Opus, yet benchmark deltas appear different.

Frontier / Code & agentic benchmarks

A reported “Frontier Code Diamond”:

  • Opus 4.8: ~half of the shown Fable/Mythos results

OpenAI “GPValve”

GPT-5.5 and Opus 4.8 numbers are contrasted, with emphasis on differences in:

  • real-world knowledge work
  • spatial reasoning

Tool/agent ability

Benchmarks covering tool use and computer use are described as improved (speaker provides percentage comparisons).

Legal/agent and complex reasoning exams

“Scarily named” benchmarks are mentioned with improved scores, including:

  • Legal Agent Benchmark
  • Humanity’s last exam
  • Terminal Bench (emphasized for agentic coding)

The speaker cites a Terminal Bench comparison (as stated):

  • ~83.4% vs 88%

What the speaker claims from firsthand testing

“Feels like massive exploration”

Across tasks, the speaker reports Fable behaves as if it is kicking off broad, deep exploration of the entire project/codebase:

  • It appears to examine many angles of every line
  • Sometimes it even covers “projects [the user] hasn’t touched in years”
  • Small tasks can feel less “small” after submission—responses are described as substantial and deep
  • For very complex, long-horizon tasks, the speaker says it doesn’t hiccup and often feels “almost too simple”

Long working time / sustained performance

  • The speaker claims it consistently takes 5+ minutes minimum
  • Rather than quick answers

A cited report (“Drip,” as referenced by the speaker) claims Fable compressed:

  • months of engineering into days for a 50M-line Ruby codebase:

  • migration in about a day vs 2 months by a team

Token efficiency & output “information density”

The speaker argues Fable is more token efficient than prior Claude models, based on:

  • high information density
  • dense information allegedly enables more progress per unit output

Downsides/quirks noted:

  • output can be hard to read
  • can be verbose
  • can be overly technical
  • may require prompts like “simplify the explanation” to make results usable

Workflow features & parallelism

A major practical unlock described is using Anthropic’s workflows feature.

Ultra Code / workflows

  • Creates a planning agent that delegates to many sub-agents in parallel
  • The speaker claims it can use up to hundreds of agents

Example:

  • “review my entire codebase” triggers sub-agents per file
  • resulting in 100+ agents running concurrently
  • token usage increases rapidly

Recommendation from the speaker:

  • start with the lowest “thinking/effort” setting, because higher effort can be overkill

Safety, guard rails, and privacy policy

Safety/guard rails

Safeguards are described as conservative:

  • subtitles mention guard rails catching “harmless requests” in <5% of sessions
  • the speaker says they did not personally encounter it

Data retention policy (Mythos-class models)

  • 30-day retention for traffic on Mythos-class models (first- and third-party services)
  • not used for training “new cloud models” or non-safety purposes (as stated)
  • logging/deletion and defenses against:
    • novel jailbreaks
    • multi-request attacks

Pricing & routing advice

Cost

  • $10 per million input tokens
  • $50 per million output tokens

The speaker calls this “incredibly expensive.”

Model routing strategy

The speaker recommends:

  • Fable for the hardest tasks
  • cheaper models (e.g., Sonnet/Haiku) for routine work

It’s also claimed to be less than half the price of Claude Mythos preview.

Release context & skepticism about “safety marketing”

The speaker speculates about Anthropic’s strategy:

  • Mythos allegedly tested internally since January, released later
  • date mentioned: June 9, 2026
  • release delay may relate to keeping a research lead while developing the next model

The speaker also raises a concern:

  • advanced models could be distilled into less-safe variants
  • if distillation is attempted, requests may “fallback” to Opus 4.8 (as stated)

Notable review / demo tests mentioned (with outcomes)

Pokemon FireRed (vision only)

  • Claims Claude Fable 5 can play through with vision alone
  • without maps/navigation aids
  • unlike earlier models that allegedly needed harnesses

Rubik’s Cube (simulation)

  • Claims it builds an interactive playable 3D Rubik’s Cube
  • can scramble
  • then solve successfully

Fluid dynamics simulation

  • Claims an interactive real-time fluid sim runs flawlessly in the browser
  • using many parallel agents and large token budgets per agent
  • speaker cites: 63 agents, 20–30k tokens each

Solar system / eclipses simulation

  • another vision/physics-style demonstration is mentioned

Practical “tutorial-like” takeaway (usage pattern)

The video emphasizes a practical workflow pattern (not a formal tutorial):

To unlock the model’s full potential, combine:

  1. Fable + workflows (parallel delegation)
  2. Fable + loops (continuous/goal-driven iterative “software factory” behavior)

The speaker frames loops/workflows as higher-level abstractions enabling software-factory automation.

Sponsors / source of publishing tool (mentioned in the video)

The video is brought to you by Here.Now:

  • a tool for enabling an agent to publish to the web
  • via a simple “install skill / copy setup” flow

Mentions:

  • private storage (not everything must be public)
  • custom URLs via a custom domain
  • “completely free right now” (as stated)

The speaker says Here.Now helped publish tests during a live stream.

Main speakers / sources

  • Primary speaker: The YouTube creator reviewing Claude Mythos 5 / Fable 5 (unnamed in subtitles)
  • Referenced sources/entities:
    • Anthropic (Claude Mythos/Fable releases, safety/privacy policies, workflows positioning)
    • OpenAI (GPValve benchmark; also “Nome Brown” blog referenced)
    • Nome Brown (OpenAI researcher): blog post claim about thinking tokens vs quality
    • Here.Now: sponsor/tool provider for publishing agent outputs

Original video