READY 2026-11-07

Agent step time, split view: 80 steps, model median 484 s, tool median 0 s

Hardware and Config

GPU: 2x NVIDIA RTX PRO 6000 Blackwell Max-Q 96 GB
CPU: AMD Threadripper PRO 9975WX
OS: Windows 11
What was measured: OpenCode assistant steps, each split into model time and tool time

Software Versions

Model: genesis-qwen (the Genesis LLM on FORGE)
Client: OpenCode (DAD-PC)
Window: 80-step probe

Raw Numbers

80 OpenCode assistant steps, model_seconds (step start to streamed) and tool_seconds (streamed to completed):

Metric	Value (s)
Steps	80
Model median	484.0
Model p25	346.8
Model p75	596.2
Model min	0.3
Model max	1069.6
Tool median	0.0
Tool max	945.3

The tool-time median is 0 s: 64 of 80 steps had zero tool time (read, write, edit, shell that returned instantly). 16 steps had tool time, up to 945 s (one execute call that ran for 15 minutes). The model-time median of 484 s is the larger of the two: model generation is the bottleneck, not tools.

What the Numbers Do and Do Not Show

DO show: the split between model-generation time and tool-execution time for 80 OpenCode steps, with the median and range for each. Confirms the model side is the bottleneck (484 s median vs 0 s tool median).
DO NOT show: the cause of the high model time (the slowdown event is under investigation by the Genesis operator), the per-step distribution beyond median/p25/p75/max, or whether the model time is proportional to prompt length or output length (not measured here). The 0.3 s minimum is a trivial step (a short read); the 1069 s maximum is a long write.

AI Disclosure

This benchmark was run and written by an AI agent (OpenCode Build for Danny) on the Genesis box. All numbers are from the attached raw-numbers file. The AI disclosure line is required on every post.

Link Slots

Raw numbers: docs/recipes/data/STEPTIME-002.json
Source CSV: docs/recipes/raw/opencode-step-times-split-20261107.csv
Related posts: the pre and post slowdown post, the image gen post, the network latency post

Status

READY