Opus 5 Beats Fable... at Half the Price?

Overall claim / positioning

• Anthropic released Claude Opus 5, positioned as a “thoughtful and proactive” model.
• It’s framed as approaching Claude Fable 5-level intelligence while costing about half the price.
• It’s described as:
  • Daily-use ready
  • More efficient than other models
  • Default on Claude Max
  • Strongest model on Claude Pro
• Cybersecurity: Opus 5 is said to remain behind Mythos 5 on cybersecurity tasks.

Benchmarks & performance analysis (why Opus 5 is competitive)

• Artificial analysis / third-party benchmarks are cited as showing Opus 5 “winning” on many metrics, often by non-trivial margins.
• Frontier Bench: Opus 5 is claimed to be ~10% better than Fable 5.
• Arc-AGI benchmark (major highlight):
  • Opus 4.8: 1.5%
  • Prior high (Soul): 7.8%
  • Opus 5: 30.2%
  • The Arc-AGI team attributes this to a new capability: the model converts tests into algebraic notation to perform advanced logical reasoning.
• Data retention constraint affecting Fable:
  • Fable isn’t shown on Arc-AGI due to data retention policy concerns (to avoid leaking details).
  • The video claims Opus 5 has no such general access retention requirements, which is framed as beneficial for users.

Coding, intelligence, and cost-to-complete comparisons

Coding & intelligence metrics

• Coding index (Artificial Analysis):
  • Opus 5 is 2nd, close to GPT-5.6-Soul Extra High (only ~0.3 points behind).
  • It beats Fable 5 by ~1.5 points and improves over Opus 4.8.
• Aent(I)tic index / intelligence index:
  • Opus 5 is again described as slightly ahead of Fable 5 and above OpenAI’s Soul model.

Price-to-performance (“cost to complete a task”)

• Opus 5 is reported as ~72 cents cheaper than Fable 5 in one cost metric.
• It’s also described as only slightly above the cost of Opus 4.8.
• However, the video notes GPT 5.6 Soul still wins overall price/performance:
  • Nearly half the price of Opus 5.
• Cursor Bench is said to match similar trends:
  • Opus 5 Max ~same score as Fable 5
  • but much cheaper (example: Fable costs $17 vs Opus $8 for the same task)
• Token pricing details for Opus 5 (per the video):
  • $5 per 1M input tokens
  • $25 per 1M output tokens
  • Same as Opus 4.8

Local “in action” tests (game + full-stack app)

Test 1: “Single-shot” Formula 1 racing game (single HTML file)

Opus 5 (best in the video’s opinion)

• Strong visuals
• No external assets
• Track/game logic works (lap timing/positioning)
• Collision works
• Grass area has an underlying track that can be revealed via prompting
• Gravel traps behave well

Fable 5

• Also produces a working game with decent visuals
• Cars are “more basic”
• Includes camera wobble; otherwise slightly less preferred than Opus 5

GPT 5.6 Soul

• Good modeling/camera look
• Issues: missing track parts and backwards assets
• Track “breaks” about halfway through (worse than Opus/Fable)

Kimmy K3 (open-weight)

• Surprisingly strong: barrier/track/positioning working

Timing & cost (F1 test)

• Opus 5 took ~46m 51s
• Estimated ~$12.99 API cost
• It was more expensive than Fable in this single run, attributed to ~5x more tokens
• Pricing accuracy was cautioned as potentially affected by subscription/tooling differences

---

Test 2: Personal finance management dashboard (full-stack)

Opus 5

• Strong UI (reviewer’s preference)
• Functional multi-page layout (separate pages per requested features)
• Front end: React + React Router
• Back end: Node/SQL (SQLite)-type database + Express

Fable 5

• Similar backend stack (React + routing choice and Node/SQLite + Express)
• UI charts criticized as sometimes “useless”
• Features work, but UI investment seems less than Opus

GPT 5.6 Soul

• Excellent UI (reviewer’s style preference)
• Routing behavior differs (changes sections rather than separate feature pages)
• Uses a more “unique” stack: Next.js + Drizzle
• Hosted with Cloudflare + Vinex
• Reviewer suggests this may be driven by prompts or platform conventions

Kimmy K3

• UI works and is “adequate,” slightly behind Opus in polish
• Routing: built its own routing library
• Major limitation: no real database; instead stores data to a JSON file via JavaScript (reviewer dislikes this)

Timing & cost (finance dashboard test)

• Fable: $30.79, 56m 55s
• Opus: $29.82, 59m 15s
• GPT models described as ~3.5x cheaper
• Reviewer concludes results are mixed, partly because only two runs were performed and cost measurement may differ from subscription pricing

Cybersecurity policy / safeguards difference (important limitation)

• Opus 5 is described as slightly less strict than Fable 5 on some cybersecurity work.
• Safeguards are similar to Opus 4.8, except for stronger guardrails on certain narrow cyber tasks.
• Claimed behavior:
  • May allow vulnerability finding in source code
  • Blocks binary-based vulnerability scanning, penetration testing, and exploit generation
• One reported test result:
  • Classifiers intervene ~85% less often than with Fable 5

Key takeaways from the video

• Opus 5 is presented as a strong “default” premium model with excellent benchmark performance, notably on Arc-AGI, and competitive coding/intelligence metrics.
• Aggregate pricing/efficiency looks favorable, but in local tests Opus 5 isn’t always cheaper (possibly due to higher token usage in some runs).
• In hands-on tasks:
  • Opus 5 is favored for UI quality (finance dashboard) and game completeness/polish (F1 HTML game)
  • Kimmy K3 is surprisingly capable but may lack robust database implementation
• Safety/cyber:
  • Opus 5 may be more permissive for source-code vulnerability discovery while still blocking more dangerous workflows.

Main speakers/sources

• Primary speaker: The YouTube narrator/reviewer (no specific name given in the subtitles)
• Referenced third parties / benchmark sources:
  • Artificial Analysis
  • Arc-AGI benchmark team