I expected Claude Code to crush its open-source rival on my 16GB GPU, but that’s not what happened - XDA

The era of expensive, cloud-only AI agents may be drawing to a close. A developer recently demonstrated that a locally run, open-source coding agent can perform complex tasks with the same competence as Anthropic's flagship Claude Code, which is sold as a $20 monthly subscription.

Abhinav Raj, a tech writer and editor-at-large, pitted the open-source stack against Claude Code to see if a $20 monthly fee was still justified. He equipped his PC with an RTX 4070 Ti Super and a Ryzen 5 7600X, running the Qwen3.6-35B-A3B model through the OpenCode v2.2.0 harness.

The Qwen3.6-35B-A3B is a mixture-of-experts model that activates only around 3 billion parameters per token. This architecture, combined with optimizations from llama.cpp, allowed Raj to run the 17.7GB model on a 16GB GPU by offloading 12 layers of experts to system RAM. The setup achieved a generation speed of roughly 65–70 tokens per second.

A complex coding challenge

To test the agents, Raj gave both systems the same job: build a Python tool that logs an Nvidia GPU's power draw, temperature, utilization, VRAM usage, and core clock every second to a CSV file. The tool needed to include a --duration option, calculate average and peak power, and save a chart of power and temperature over time as a PNG. It also had to perform a 30-second self-test.

The local agent, Qwen3.6-35B-A3B via OpenCode, completed the task in 3 minutes and 44 seconds without any human intervention. It found the nvidia-smi command on its own, wrote the script, revised it, and tested it. The utility it produced worked flawlessly, and its reported average and peak power draw of 27.32W and 74.83W matched the raw CSV data and the official nvidia-smi output.

Claude Code's edge is polish, not utility

Claude Code, powered by the Claude Sonnet 5.5 model, completed the same job in 2 minutes and 42 seconds. It was about a minute faster from start to finish and ran its 30-second test three times to fix visual problems with chart labels. The difference was in the polish: Claude double-checked its math, mentioned that the chart needed the matplotlib library, saved the log after every row, and split power and temperature into separate panels.

Both agents delivered a working Python utility with correct numbers, so the gap between them comes down to effort and cost. Raj noted that the extra effort Claude put into the polish would be worth it for most users. However, the local agent did the same job for a fraction of a cent in electricity, while the Claude session cost about $0.37.

The local agent's performance was so strong that it begs the question of what the $20 monthly subscription is buying. For a workflow that does not require flagship model capabilities, the open-source alternative offers a compelling, cost-effective alternative.