Anthropic's Claude Opus 5 Tops Agentic Benchmarks at 20% Lower Cost

Claude Opus 5 just landed, and it immediately reshuffled the leaderboard for agentic knowledge work. Anthropic's new model takes the top spot on the AA-Briefcase benchmark, Artificial Analysis's proprietary evaluation for long-horizon, real-world office tasks, outperforming Claude Fable 5 by 146 Elo points while costing 20% less per task. That's a meaningful combination: more capable and cheaper, at a tier that was already considered frontier.

What AA-Briefcase actually measures

AA-Briefcase is not a multiple-choice quiz. It evaluates models across four multi-week knowledge work projects, comprising thousands of input files and 91 tasks in total, covering fields such as data science, product management, and corporate strategy. Each task produces a real deliverable, think Excel spreadsheets, PowerPoint decks, research memos, and PDFs, graded across three dimensions:

• Rubric pass rate: binary pass/fail checks on whether the model followed instructions, found requirements buried across source files, and reached correct conclusions
• Analytical quality: pairwise head-to-head comparison against another model's submission for rigor and depth
• Presentation quality: pairwise comparison for professional formatting and clarity

The combined score is expressed as an Elo rating, the same system used in chess rankings, where a 100-point gap represents a significant and consistent performance difference.

The numbers that matter

Claude Opus 5 is available today and comes close to the frontier intelligence of Claude Fable 5 at half the price. On AA-Briefcase specifically, the five effort settings tell a nuanced story:

• Max effort: 1720 Elo, $17.79 per task (20% cheaper than Fable 5's $22.30)
• Xhigh effort: 1693 Elo, $14.26 per task (64% of Fable 5's cost)
• High effort: 1606 Elo, $10.41 per task (47% of Fable 5's cost, still beats Fable 5 by 32 Elo)