Intel’s Xeon 6 completed up to 1.66 times as many agent tasks as a 5th Gen AMD EPYC instance in a 60-minute run, and 4.24 to 4.57 times as many as an Arm v9.2A instance across the three verticals, according to an Intel engineering post. The tasks come from Terminal-Bench 2.0, and no model took part: Intel replays a recorded Claude trajectory through a local proxy, so no instance runs inference or touches the network.
Three 48-vCPU AWS instances each ran 24 sandboxes at a time for an hour
The Xeon 6 machine is an m8id.12xlarge, the EPYC one an m8a.12xlarge and the Arm one an m9gd.12xlarge. Each has 48 vCPUs and 185 GB of memory and ran Ubuntu 26.04 on the 7.0.0-1006-aws kernel. Intel used Harbor 0.16.1 with the terminus-2 agent, kept 24 sandboxes running at a time with one pinned to each core, and refilled slots as tasks finished. The metric is tasks completed per 60 minutes across the whole box, with the verifier on, and all three instances finished every task without failures. Intel tested in September 2026.
It splits the same workload into three verticals, and the ratios barely move between them:
| Vertical | Xeon 6 vs 5th Gen EPYC | Xeon 6 vs Arm v9.2A |
|---|---|---|
| Healthcare | 1.66 times | 4.5 times |
| Banking and financial services | 1.64 times | 4.57 times |
| Manufacturing | 1.58 times | 4.24 times |
Replaying a recorded trace times the machine around the model, not the model
Since the model is a recording, the clock runs on everything an agent does on its host: starting containers and shell commands, orchestration, memory and storage traffic. That is where Intel puts the bottleneck. Its authors write that deployments often find “the limiting factor is not model generation speed”, with the slowdowns appearing in task orchestration, tool execution, memory access, storage and runtime management. The 1.66 times figure appears to measure that host-side part of an agent job. Tokens per second, GPU throughput and local inference sit outside it.
The post gives its results as ratios rather than task counts, and it does not price the three instances, so the figures do not show what an hour of agent work costs on each. Intel ran all three tests itself.