APEX-Agents is Mercor’s frontier benchmark for long-horizon cross-application tasks in professional services. It tests agents' ability to do high-value work in investment banking, management…

One caveat. Anthropic notes the new classifier flags more benign coding requests, and we see that in the shifted tool mix. Small output-type samples are noisy too (slides n=6). But on the metric that…

One caveat. Anthropic notes the new classifier flags more benign coding requests, and we see that in the shifted tool mix. Small output-type samples are noisy too (slides n=6). But on the metric that…

One caveat. Anthropic notes the new classifier flags more benign coding requests, and we see that in the shifted tool mix. Small output-type samples are noisy too (slides n=6). But on the metric that…

Two APEX-Agents domains demonstrated big improvements from Sonnet 4.6. For Investment Banking Analyst tasks, Sonnet 5 moved up from #18 to #8 (+11.2pts) mean score. For Management Consultant tasks,…

APEX-Agents domain breakdown for Claude Fable 5 (Max), Pass@1: Corporate Law: 40.9% (1st) Investment Banking: 47.7% (2nd) Management Consulting: 46.4% (2nd) With 4 runs, Fable solved 246 of 480…

APEX-Agents domain breakdown for Claude Fable 5 (Max), Pass@1: Corporate Law: 40.9% (1st) Investment Banking: 47.7% (2nd) Management Consulting: 46.4% (2nd) With 4 runs, Fable solved 246 of 480…

With 4 runs, @claudeai Opus 4.8 solved 251 of the 480 tasks in APEX-Agents, 37 more than Opus 4.7. Opus 4.8 is more efficient than other models, using 36% fewer tokens than GPT 5.5 and 20% fewer than…

The tradeoff is clear: Gemini 3.5 Flash wins on raw task completion, but it gets there by using significantly more compute. Whether that tradeoff makes sense depends on what you're optimizing for.…