TB-Rewind
TB-Rewind is our newest agentic test. It replays human reference solutions for 18 Terminal-Bench tasks, 5 from Terminal-Bench 2.0 and 13 from 4.0, in frozen containers with no LLM in the loop. Many use Terminal-Bench 2.0, but after profiling dozens of systems, we opted for a smaller set of workloads from TB 2.0 and 4.0 that were more consistent.
We show this in runs per hour, so higher is better. One catch is that this is a speed test and not a capacity test. Core count barely moves it, because only one task runs at a time.

The EPYC 9996 completes 38.8 suite runs per hour, the most here, but only 2% ahead of Vera at 37.9. The EPYC 9965 is at 32.3, Grace at 30.5, and the Xeon 6980P at 21.3. TB-Rewind runs one task at a time, so single-thread speed matters most here, which bodes very well for the Zen 6 microarchitecture given the competition.
Here is each task.

The standout is probably the Vera qemu-startup number.
That is hard to compare across 18 different scales, so here it is relative to the EPYC 9996.

The EPYC 9996 is the fastest chip on 9 of 18 tasks. Vera beats it on most of the rest, so the suite result is close to a tie between the two, but this is probably one of the most interesting benchmarks.


