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

하이라이트

• Claude Opus 5 scores 30.2% on the Arc AGI benchmark by converting visual reasoning tests into algebraic notation.

• Artificial Analysis ranks Opus 5 second on its coding index, beating Claude Fable 5 by 1.5 points while trailing GPT 5.6 Sol Extra High by 0.3 points.

• Unlike Fable 5, Opus 5 does not impose strict data retention requirements for general API access.

• Opus 5 pricing stands at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8.

• Anthropic security classifiers intervene 85% less often on Opus 5 than on Fable 5, allowing source code vulnerability scanning while still blocking exploit generation.

타임라인

1. 00:00 - 01:31

   Opus 5 Benchmark Performance and Arc AGI Breakthrough

   Benchmark evaluations position Opus 5 near the intelligence threshold of Fable 5 while operating at half the price. The model converts Arc AGI spatial puzzles into algebraic notation to execute advanced logic, demonstrating a mechanism unseen in previous generations. General access to Opus 5 removes the data retention restrictions required by Fable 5, allowing benchmark maintainers and enterprise users to run evaluations without data exposure risks.

   • Opus 5 serves as the new default model on Claude Max and the top model on Claude Pro plans.
   • Frontier Bench results place Opus 5 nearly 10% ahead of Fable 5.
   • Arc AGI evaluation scores show Opus 5 reaching 30.2%, up from Opus 4.8 at 1.5% and Sol at 7.8%.
2. 01:31 - 03:26

   Third-Party Index Ratings and Pricing Dynamics

   Evaluations from Artificial Analysis place Opus 5 behind GPT 5.6 Sol Extra High by 0.3 points on the coding index while outperforming Fable 5 by 1.5 points. Token pricing remains fixed at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8 rates. Fable 5 retains an edge on long-horizon, complex tasks, but Opus 5 covers approximately 95% of standard development workloads.

   • Opus 5 captures first place on both the Artificial Analysis Agentic Index and Intelligence Index.
   • Cursor Bench evaluates Opus 5 Max and Fable 5 at near-identical output quality, with task costs of $8 and $17 respectively.
   • GPT 5.6 Sol maintains the overall price-to-performance advantage across third-party index evaluations.
3. 03:26 - 05:56

   3D Game Generation Benchmark: 3.js and HTML

   Testing across models required rendering a 3D racing game using 3.js inside a self-contained HTML file. Opus 5 delivered superior vehicle modeling, collision handling, and lap tracking, outperforming Fable 5's simplified geometry and GPT 5.6 Sol's broken track rendering. However, Opus 5 ran for 46 minutes and 51 seconds, generating excess output tokens that drove API costs above Fable 5.

   • Opus 5 generated a functional 3D Formula 1 racing game within a single HTML file without external assets.
   • Single-shot generations across Fable 5, GPT 5.6 Sol, and Kimi K3 produced playable track mechanics.
   • Opus 5 consumed five times more tokens than Fable 5, resulting in a higher task cost of $12.99 versus Fable's lower execution spend.
4. 05:56 - 09:26

   Full-Stack Web Development and Architecture Selection

   Full-stack dashboard creation tests revealed architectural choices across models. Opus 5 selected React Router and Node SQLite for a persistent backend structure with complete page routing. Fable 5 wrote a custom inline routing handler rather than importing React Router, while Kimi K3 failed to integrate an SQL database. Task pricing for the application run showed GPT 5.6 Sol operating roughly 3.5 times cheaper than Anthropic models.

   • Opus 5 constructed a full-stack personal finance application using React, React Router, Express, and Node SQLite.
   • GPT 5.6 Sol generated a comparable user interface using NextJS, Drizzle, Cloudflare, and Vinext.
   • Kimi K3 omitted a true database driver, substituting an in-memory JavaScript JSON file store.
5. 09:26 - 10:14

   Cybersecurity Safeguards and Market Positioning

   Refined safety guardrails allow Opus 5 to analyze source code for vulnerabilities without triggering false positives. Binary exploitation, active penetration testing, and automated exploit generation remain prohibited. The broader market landscape exhibits price reductions led by GPT 5.6, intelligence gains from open-weight models like Kimi K3, and top-tier reasoning performance from Anthropic.

   • Opus 5 safety classifiers intervene 85% less frequently than Fable 5 guardrails.
   • Source code vulnerability scanning is permitted, while binary-based exploitation and penetration testing remain blocked.
   • Open-weight models like Kimi K3 continue closing performance gaps against proprietary frontiers.