Five coding agents on one local Qwen: Claude Code finishes in 15 min while Codex burns 4 hours and aborts

seti_at_home · reddit · 2026-11-13

The author ran Claude Code, OpenCode, GitHub Copilot Desktop, Pi, and Codex against the same local model—Qwen3.8 Flash Next (Unsloth UD-Q4KXL) served by Halogen Flash Server 0.20.4—on an AMD Strix Halo machine (Ryzen AI Max+ 395, 128GB unified memory), all building the same looping pixel-art Halloween diorama, with 30-second clips of each result.

Server-measured stats (one run per client, not a controlled benchmark):

| Agent | Elapsed | Result | Requests | Output tokens | Decode tok/s | Cache reuse |

|---|---|---|---|---|---|---|

| Claude Code | 14m42s | Final answer | 15 | 31,373 | 38.1 | 95.6% |

| OpenCode | 36m30s | Final response | 54 | 77,832 | 39.3 | 98.3% |

| Copilot Desktop | 58m30s | Final response | 29 | 54,393 | 20.2 | 85.4% |

| Pi | 1h18m | Last request before memory stall | 108 | 135,469 | 38.7 | 98.6% |

| Codex | 4h11m | Turn aborted | 355 | 402,803 | 37.0 | 96.9% |

Claude Code reached a final answer in 15 minutes with 31K output tokens; OpenCode used 2.5x tokens; Copilot's three full prompt rereads added 10 minutes of prefill; Pi hit a tool loop until the server stalled on memory; Codex made 355 requests with a peak prompt of 229K tokens before the turn was aborted. Clients used different endpoints (Chat Completions, Anthropic Messages, Responses) with their own tool definitions and history management. Output quality was not scored; the video lets readers compare themselves.

Original post →

More from coding & agent

coding & agent channel →