Coding Agent Showdown: Claude Code is 4x Slower With No Quality Gain
xquarx · reddit · 2026-07-27
The author benchmarked DeepSeek V4 Flash across Claude Code, OpenCode, and Pi, finding that while code quality remained identical, the time and token consumption varied wildly.
- Core Finding: Given the same underlying model, the harness doesn't change code quality but drastically impacts efficiency. Claude Code took nearly 4x longer than the fastest tool.
- Behavioral Differences: Each framework took radically different paths. Pi reasons directly, OpenCode delegates tasks, while Claude Code loves exploring the codebase (perhaps excessively).
- The Bottleneck: System prompts and tool call structures play a massive role in how efficiently the agent operates. Detailed charts and raw per-run data are公开 for the community to inspect.
More from coding & agent
- Planr adds a live map to show what each agent is doing and what is blocked — kevinkern · 2026-07-27
- Planr adds a live workflow map to track multiple agents, blockers and dependencies — kevinkern · 2026-07-27
- Long-running agents hit race conditions after immutable event logs — neal_lathia · 2026-07-27
- An autonomous engineering loop shipped 7 production PRs while the developer was away — bestofdesp · 2026-07-27
- An AI coding agent is using computer-use tools to test the AI tool it’s building — carsonfarmer · 2026-07-27
- Codex bypasses AllTrails CAPTCHA, automates browser scraping from a simple hike query — yacineMTB · 2026-07-27