Same DeepSeek Model, Two Shells: CLI Won Speed and Precision, GUI Won Polish
Which shell wraps a paid model changes the bill, the edit style, and how much unrequested flourish lands in the output, even with the model and prompt held fixed. Token counts alone don't predict cost — a 206K-token run at 99% cache hit billed less than a 58K-token run.
Two shells drove the identical DeepSeek model across three coding tasks with word-for-word identical prompts: Claude Code over the API from a terminal, and DeepSeek's official open-source desktop harness DSH. The whole experiment cost ¥0.36, and every round passed on the first try with zero errors on both sides.
The pattern repeated each round. The CLI finished all three tasks in 74 seconds total versus 117 seconds for the GUI, cost about ¥0.16 versus ¥0.20, and edited code surgically — 9 lines added and 8 removed for the dark-mode color change, where the GUI rewrote larger blocks. The GUI's output surprised more: 11,134 output tokens against the CLI's 4,400 in round one, adding entrance animations, favicons, keyboard focus styles, and numbered project cards that no prompt requested.
Cache behavior split too. DSH held 95–99% cache hit rates throughout; Claude Code started at 68.2% in round one before climbing to 99% in later rounds. The daily driver went to DSH anyway. GUI simplicity, steadier cache hits, and desktop operation history outweighed the CLI's speed, cost, and edit-precision wins, which left server-side and batch work to the CLI.
Same model, same prompt, but different outputs: the CLI delivered the literal requirement, while the GUI reached for extra design polish; the harness encodes a behavioral default (engineer vs publisher) that the model itself doesn't.
Round two is where token accounting misleads: a 3.5x token load cost less because of cache warmth, so agent cost forecasts built on raw token counts will misprice both tools.
Diff hygiene for a dark-mode switch touched only 17 lines in the CLI run. That matters more as codebases age than raw speed; edit precision is the quality that prevents regressions.
Choosing DSH as the daily driver shows the decision isn't purely rational cost-per-task; daily friction, history browsing, and mental load override a 20% cost advantage.