Anthropic、エージェント型知識作業で新リーダー「Claude Opus 5」を発表

See model page

Claude Opus 5 is the new leader on our agentic knowledge work benchmark, AA-Briefcase, outperforming Claude Fable 5 by nearly 150 Elo while reducing Cost per Task by 20%

Anthropic has released Claude Opus 5, the new leader on the Artificial Analysis Intelligence Index, and also the #1 model on our proprietary agentic knowledge work benchmark, AA-Briefcase. At max effort, Claude Opus 5 scores an AA-Briefcase Elo of 1720, 146 points ahead of Claude Fable 5 (1574). Its xhigh (1693) and high (1606) variants also outperform Fable 5 while being more cost efficient

We benchmarked all five effort settings ahead of release. Claude Opus 5 (max) costs $17.79 per task, a 20% reduction from Claude Fable 5 ($22.30). The xhigh and high variants offer stronger cost-efficiency tradeoffs, both scoring above Fable 5 while costing 64% ($14.26) and 47% ($10.41) as much respectively

AA-Briefcase is our proprietary benchmark for agentic knowledge work, evaluating models on realistic, private tasks spanning thousands of input files and requiring deliverables such as research reports, presentations, and spreadsheets. Performance is measured across correctness, analytical quality, and presentation quality, and combined into a single metric, the AA-Briefcase Elo

Key AA-Briefcase results across all 5 Claude Opus 5 effort settings:

➤ The clear leader in agentic knowledge work: Claude Opus 5’s top three effort settings (max, xhigh, and high) take the top three positions on AA-Briefcase. Combined with Claude Fable 5, Sonnet 5 (max), and Opus 4.8 (max), Anthropic now holds a large majority of the top ten spots on the AA-Briefcase leaderboard. Claude Opus 5 scores 1470 Elo at medium effort, just behind GPT-5.6 Sol (max, 1505), and 1223 at low effort, just below GLM-5.2 (max, 1254)

➤ Leads in objective criteria and analytical quality, but not presentation: AA-Briefcase measures performance across objective rubric criteria, analytical quality, and presentation quality. Claude Opus 5’s gains are driven primarily by rubric pass rate and analytical quality. At max effort it achieves an Analytical Quality Elo of 2016, nearly 300 Elo ahead of Claude Fable 5. Compared to Opus 4.8, presentation quality sees a smaller improvement, with a Presentation Elo of 1628, still ~40 Elo behind GPT-5.6 Sol (max, 1666)

➤ Half the cost for Fable-level knowledge work capabilities: Claude Opus 5 offers a wide range of intelligence-cost tradeoffs across its five effort settings. Max effort costs $17.79 per task, around 20% less than Claude Fable 5 ($22.30) while achieving a higher AA-Briefcase Elo. Notably, Claude Opus 5 with high effort outperforms Claude Fable 5 by 32 Elo while costing $10.41 per task, less than half the price. Claude Opus 5 retains Opus 4.8’s pricing of $5/$25 per million input/output tokens, with a 25% premium for cache writes and a 90% discount for cache hits

➤ Over 25 minutes per task for frontier performance: Claude Opus 5’s top three effort settings all average more than 25 minutes per AA-Briefcase task (36.2, 34.3, and 25.7 minutes for max, xhigh, and high). At max effort this is roughly 50% longer than Opus 4.8 (24.1 minutes), primarily due to an increase in the number of turns. Claude Opus 5 averages 103, 91, and 76 turns per task across its top three effort levels, compared to 55 for Opus 4.8 (max)

For full AA-Briefcase results, see https://artificialanalysis.ai/evaluations/aa-briefcase