Repository navigation
guides: DeepSeek Search and Fetch free window ends 2026-11-01 - #146
Conversation
Document the 2026-11-01 00:00 Beijing cutoff for server-side Search and Fetch in Harness, Claude Code, and Codex, with the standalone API rates and a pre-cutoff log checklist. Link it from the API platforms guide. Co-authored-by: Liu <ljq258@gmail.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 83f4cbcd8b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| 1. The model-facing Harness tool `web_search` takes `queries`, from 1 to 4. The web subsystem doc says each query is fanned out into its own search. One tool call can become several searches. A one-item array is one search. | ||
| 2. The shipped Messages search provider id is `deepseek-official`, package `@deepseek-ai/dsh-web-search-deepseek`. The README says one search is a full Messages model turn, with up to `maxUses` server-side web searches inside that request. The default is **5**. If all 5 succeed and each one is a billable Search request, the Search fee is `5 / 1000 × unit price`. The README does not say every search hits 5, and it does not say every one returns a non-empty result. That path also spends model tokens. | ||
|
|
||
| The 4 on `queries` and the 5 on `maxUses` are two caps. Multiplying them into 20 successful requests is not in the docs. The bill follows the successful-request count. Log that count before the cutoff. |
There was a problem hiding this comment.
At the pinned Harness revision, packages/web/tool-web/src/search.ts::runSearchQueries calls ctx.web.search once for each of the up-to-four queries, while the web-search-deepseek README allows up to maxUses = 5 server-side searches in each resulting Messages request. The documented worst case for deepseek-official is therefore 4 × 5 = 20 server-side searches per model-facing call; saying there is no basis for multiplying these limits understates the post-cutoff cost ceiling by up to 4×. Keep the caveat that actual billing depends on successful searches, but present 20 as the theoretical maximum.
Useful? React with 👍 / 👎.
What changed
Bilingual howto guide for the DeepSeek Search and Fetch billing cutoff. Frontmatter
draft: falseso it publishes when this PR merges. The PR itself is a draft for Bai to review.src/content/guides/deepseek-search-fetch-free-window.mdsrc/content/guides-en/deepseek-search-fetch-free-window.mdGuides are picked up by the content collection. Build lists both routes, both RSS feeds, and the sitemap. Related cards on the new page are Harness intro, API platforms, and the vision guide.
The article states there was no live API call and no console screenshot. Counts in the budget section are arithmetic from the published per-1,000 rate, not measured usage.
Evidence opened 2026-10-23 (Asia/Shanghai)
resultsarray, or Fetch page content. Failures, empty results, and timeouts are not billed.countdoes not change the per-request price. Retry guidance: wait and retry 408, 429, 5xx; do not retry other 4xx unchanged; 402 means top up first. QPS: Search 100, Fetch 25, plus an unpublished per-domain Fetch limit.deepseek-flash.models.jsonsetsweb_search_tool_typetotext. No documented off value.web_searchruns fordeepseek-v4-proand is ignored fordeepseek-flash.max_tool_callsis ignored.687f282e8625c6b8d11ff58211a90524be2e4317(2026-10-17, 0.2.1-rc.2):packages/web/web-deepseek-api/README.mdandpackages/web/web-search-deepseek/README.md. Shipped selection is Messages search plus local HTTP fetch. Use the new Web Search API / 启用新版 Web 搜索 API selectsdeepseek-search-apianddeepseek-fetch-api(direct/v1/tools/searchand/fetch, no automatic retry).deepseek-officialallows up tomaxUses5 server-side searches per Messages request. Env namesDSH_WEB_SEARCH_PROVIDERandDSH_WEB_FETCH_PROVIDERare inpackages/web/web/src/index.ts, not on the published subsystem page.Open questions for Bai
/v1/tools/*call billed now? The pricing note names products, not plugin idsdeepseek-search-api/deepseek-fetch-api. The article tells readers to budget as if it may already be billed.deepseek-flashignoresweb_search. The Claude Code page does not say Web Search is Pro-only, and its sample usesdeepseek-flash. Do flash sessions in Claude Code still emit Search calls?@deepseek-ai/dsh-web-fetch-httpexists. Whether it is the shipped selection is marked unknown.