跪拜 Guibai
← All articles
DeepSeek · Claude

Same DeepSeek Model, Two Shells: CLI Won Speed and Precision, GUI Won Polish

By 逻辑与温度 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

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.

Summary

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.

Takeaways
— Both shells ran deepseek-flash with identical word-for-word prompts across three rounds and passed every round on the first try with zero errors.
— Three rounds cost ¥0.36 total: roughly ¥0.16 for Claude Code and ¥0.20 for DSH.
— Claude Code completed all three tasks in 74 seconds versus 117 seconds for DSH.
— In round one, DSH emitted 11,134 output tokens (430 lines) against the CLI's ~4,400 (323 lines), spending 58 seconds to Claude's 28.
— A dark-mode color change in round two got a 9-add/8-delete surgical diff from the CLI, while DSH made broader edits of +27/−24 lines.
— The 206K token run in round two billed less (~¥0.044) than DSH's 58K-token run (~¥0.056) because 99% hit the prompt cache.
— DSH held 95–99% cache hit rates in all rounds; Claude Code started at 68.2% in round one.
— For the progress-bar animation, the CLI chose transform: scaleX on the compositor layer to avoid reflow, while DSH used a width-based CSS animation.
— DSH became the daily driver despite the CLI's wins, on GUI simplicity, steadier cache hits, and desktop operation history.
Conclusions

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.

Concepts & terms
DSH (DeepSeek Harness)
DeepSeek's official open-source desktop agent tool with a graphical interface; the underlying model is configurable, so the harness is just a shell around whatever model you point at it.
Prompt cache hit rate
The share of tokens in an agent run served from a cached prompt prefix rather than freshly generated; a warm cache keeps cost flat even when total token counts soar across many small requests.
Compositor-layer animation
Animating transform (e.g., scaleX) moves work to the GPU compositor without triggering layout reflow, unlike animating width, which recalculates layout every frame. The difference is invisible on small pages but matters as animations multiply.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗