dsh-cost-estimate: Estimating Token Usage and Cost Before Answering

Introduction¶

In the DeepSeek Harness (DSH) Web chat interface, when sending large requests, it is usually not immediately clear what the approximate cost of the call will be. The dsh-cost-estimate plugin inserts an inline notification in the chat stream before the response begins, estimating the input/output token range and cost range; after the response ends, the same line automatically updates to the actual usage, cost, and cache hit rate.

The difference from existing community plugins (such as dsh-session-cost) is that this plugin focuses on pre-estimation—because output length is inherently unpredictable, it provides ranges and continuously calibrates using actual historical values within the session.

Core Features¶

Pre-estimation: Before the model starts generating, it first displays the estimated input tokens, output token range, and cost range.
Post-response reconciliation: After the response ends (after receiving the provider’s actual usage), the same line switches to the actual tokens and cost, and shows the cache hit rate.
Threshold triggering: By default, it displays only when estimated input >= 8000 tokens or estimated cost >= ¥0.01; small requests remain silent, while large requests are always flagged.
Peak/off-peak pricing: Supports deepseek-v4-flash / v4-pro, automatically selecting the peak/off-peak pricing tier based on the current time.
Input-side accuracy: Prioritizes anchoring to the actual usage returned by each request, using CJK-aware heuristic counting as an interim fallback.
Output-side calibration: Uses a moving average to continuously refine subsequent estimates.
Pure client-side: Calculation and rendering happen in the Web client, with zero intrusion into session logs.
Multi-step support: Tool-call loops are annotated, and actual costs are accumulated.

Installation and Activation¶

dsh plugin --profile web add dsh-cost-estimate

After installation, restart dsh web and force-refresh the browser (Ctrl+Shift+R) to load the new plugin.

Notes:
* This command depends on pnpm (npm install -g pnpm).
* If Windows paths contain spaces, use a junction or a path without spaces.
* Estimates are for reference only; actual DeepSeek billing takes precedence.

Configuration¶

The following keys are supported under config in cordis.patch.yml:

minInputTokens: default 8000. Display only when the estimated input exceeds this token count.
minCostCny: default 0.01. Display only when the estimated cost exceeds this CNY amount.
defaultModel: default deepseek-v4-flash. The fallback model before the first usage anchor.
headerTokensEstimate: default 6000. The heuristic token count for the system prompt + tool schema before the first usage anchor.
defaultCacheHitRatio: default 0.5. The assumed cache hit rate before the first usage anchor.

Example configuration:

- id: cost-estimate
  name: dsh-cost-estimate
  config:
    minInputTokens: 20000
    minCostCny: 0.05

Effect Example¶

After sending a large question, before the response:

预估：输入约 9.9K tok · 输出 2.2K–7.3K tok · 费用约 ¥0.0096–¥0.0198（v4-flash）

After the response ends, the same line updates to:

实际：输入 9.1K tok · 输出 1.1K tok · 费用 ¥0.02（缓存命中 87% · v4-flash）

Notes¶

The input side is accurate after anchoring; the output side is a range estimate.
The plugin runs with the permissions of the current dsh process; check the source code and license before installing.

Summary¶

dsh-cost-estimate is suitable for developers who need fine-grained cost control and evaluation of large model call overhead during development and debugging. It implements real-time monitoring and estimation of DeepSeek API costs on the client side.

Plugin directory: https://www.skillhub.cn/plugins/Yvesgao/dsh-cost-estimate
Source repository: https://github.com/Yvesgao/dsh-cost-estimate