Please wait while we prepare your content.
One deep-dive report per model on the bench. Same five prompts, same rubric, graded blind against Claude — every report shows the exact prompt so you can rerun it on your own data.
Expensive, boring, correct
Would use again
Steady across every task
Fast baseline. Weak on agentic runs
Slow but stubborn
Best cost-per-pass on the bench
Cheap. Drops details under pressure
Great prose, unreliable structure
Cold-start latency ruins it for interactive use
Fast and cheap. Fails anything subtle