Fable is Mythos, and it is really good.

Technological concepts / product features / analysis (key points)

Model access & “safeguards”

• The widely available model is Fable 5, not Mythos 5.
• Fable 5 is positioned as “still Mythos,” but with front-loaded safeguards that:
  • refuse or block more requests in certain scenarios
  • reduce performance on some benchmarks
• Despite restrictions, the speaker reports heavy practical usage (sub accounts + inference).

Compute + cost realities (inference and session limits)

• Reported spending on inference over ~24 hours is substantial: ~$600–$2,000, plus team activity.
• 5-hour session limits are hit quickly; the speaker describes burning through limits on multiple accounts simultaneously.
• A key pain point: after hitting limits, switching to usage-based billing led to unexpectedly fast spend (claimed $100 in ~8 minutes).
• Rough estimate provided:
  • one 5-hour session ≈ 25% of weekly limit
  • after maxing a session, continuation is only “a few times per week” (max described as “four times in a week”)
• Pricing stated:
  • $10 per 1M input tokens
  • $50 per 1M output tokens
• Overall implication: Fable/Mythos can burn through limits and cost quickly, and the speaker expects future tiering/subsidization as demand patterns become clearer.

Benchmark reporting methodology and “Mythos 5 vs Fable 5” mapping

• On some charts, labels appear as “Mythos 5 / Fable 5” because:
  • some benchmark questions get refused by Fable
  • refused questions plummet scores
• Example impact:
  • Terminal Bench: large drop (e.g., “20 point drop”) attributed to blocked requests under safeguards.

Coding benchmark outcomes

swe-bench pro

• The speaker frames it as partly flawed/weak now because much of it involves:
  • existing PR text + metadata
  • the model “recreating” details
• Fable 5 strongly outperforms (contrasted with Opus models).
• Claim: the model can “recreate those PRs,” implying strong training/data coverage.

Frontier Code (general / harder / diamond tiers)

• Discussion focuses on the hardest subset (e.g., diamond tier = 50 hardest).
• Skepticism about Frontier Code scoring behavior:
  • charts allegedly show inconsistent performance across reasoning settings
  • suspected “random generator” pattern
• A more plausible internal/early dataset (“Deep SWE”) is highlighted:
  • Fable XHigh ≈ GPT 5.5
  • better value (fewer tokens spent) even if price is higher
• Token efficiency point:
  • it may be a better deal per task because it uses fewer tokens despite being “more expensive.”

Knowledge work + vision / spatial reasoning

• Vision is described as meaningfully improved versus the “Opus line.”
• The speaker claims it may lead GPT in vision (and possibly beat Gemini 3.1 Pro, per their summary), with emphasis on spatial reasoning.
• Weaker performance reported on Skatebench:
  • Google reportedly dominates (e.g., “98% often” cited)
  • Fable improves relative to earlier versions but still lags
  • the speaker says it tries to avoid reasoning on parts of the task.

“Fallback” / refusal routing and safety tradeoffs

• Safety interventions aren’t only refusals:
  • for sensitive domains, the system may route to another model (e.g., Opus 4.8)
  • or apply efficiency/prompt-steering/PEFT-like changes that reduce effectiveness
• Privacy/safety measurement section:
  • fallback rate described as low (“fewer than 5% sessions”)
  • but the speaker claims benchmark evidence suggests ~8% task-level fallback, especially on scientific questions
  • the model may score very low on some hard exam-like questions when constrained
• Concern emphasized:
  • users may pay full price for an effectively “dumber” model without clear disclosure about when downgrades occur.

Artificial Analysis benchmark claim

• The speaker claims the model is best-in-class / smartest on that benchmark:
  • including a “~5 point lead over 5.5”
• Still, they stress that fallback/denial behavior affects real implications.

UI + template quality for code-generated interfaces

• Strong opinion: UI output quality is improved and looks less template-y than earlier Opus experiences.
• Hypothesis: improved templates or better design “instinct.”

Real-world coding / porting successes (not just benchmarks)

Reported tasks include:

• Modernizing a ~15,000-line codebase into a working version via iterative debugging using:
  • error screenshots
  • corrections
• Architecture modernization moves, including examples like:
  • Prisma → (TanStack Start, Convex, Clerk)-style modernization
  • some choices described as “auto-selected” by the model based on prompts/system instructions
• Workflow debugging/diagnosis tasks, including edge cases involving auth/UI
• Cases where the model breaks core functionality in more complex UI flows (progress mixed).

Software creation demos

• The team reportedly produced:
  • a terminal-based 2.5D adventure via prompts (claimed working interactions)
  • a Rust port of T3 code into a terminal UI with clickable thread handling
  • an “Addie made a Minecraft clone” claim: fully generated assets/textures without image generation (inferred programmatic generation)

Fuzzers / test generation

• The speaker describes prompting the model to write fuzzers and test beds to find issues (e.g., “Lakebed”).
• They report finding concrete bugs during database/architecture overhauls.
• Overall argument: the model is more trustworthy/more capable at doing this end-to-end (though not perfect).

Data retention / privacy policy

• Major operational constraint:
  • Fable 5 requires 30-day retention for all traffic
• Even with Anthropic setups that claim not to save data, the speaker warns this policy implies data is stored when using Fable 5.
• Consequence: the speaker warns it may violate some org policies/laws for real-world use cases.

How to use in day-to-day engineering (practical guide statements)

Recommended workflow principles:

• Treat AI as first-line defense for bug detection and feedback.
• Test the hardest UI flows.
• Have AI create screen recordings of PRs for review before humans look.
• Humans should focus on PR review after AI has analyzed.
• Use agents/workflows to self-prompt “most of the time”, with caution since agent prompting can be low quality.

Strategic stance:

• Let AI explore broader options:
  • give vague goals
  • synthesize, test, validate, then summarize results
• Push harder problems because the ceiling is meaningfully higher.

Economic argument

• The speaker claims software economics shifted:
  • spend AI budget on many PRs so that some become mergeable
  • avoid paying high costs repeatedly to make a single PR work
• Additional advice:
  • temporarily “squeeze usage” within current subscriptions due to limited capacity and fast market changes
  • expect pricing competition and improvements.

---

Reviews / tutorial / guide content explicitly present

• Benchmark walkthrough + critique, especially:
  • Terminal Bench
  • SWE-bench pro
  • Frontier Code
  • Skatebench
  • Artificial Analysis
  • fallback behavior
• Usage guide / process recommendations for engineering teams:
  • AI as first-line QA
  • review pipeline changes
  • agent prompting guidance
  • fuzzers/test beds
• Practical build notes from porting/modernizing real codebases:
  • iterative debugging
  • architecture migration examples
• Operational cautions:
  • cost traps from usage billing
  • refusal/safety downgrades
  • 30-day data retention requirements

---

Main speakers / sources

• Main speaker: the video narrator/reviewer (speaks in first person throughout; no explicit name given in subtitles).
• Referenced sources/quotes:
  • Walden (cited as a starting point for AI usage in engineering workflows)
  • Simon (another tester mentioned in passing)
• Bench providers/platforms mentioned:
  • GitHub / GitHub Actions
  • Blacksmith sponsor
  • benchmark suites: swe-bench pro, Terminal Bench, Frontier Codebench, Skatebench, Artificial Analysis / intelligence index