Field noteNov 20, 20266 min read1 source

DeepSeek Harness v0.2.2-rc.5 uses server-side search with no per-search charge for DeepSeek models

In one minute

DeepSeek Harness v0.2.2-rc.5 makes search routing depend on the selected model family. DeepSeek’s release notes say DeepSeek models use server-side search with no per-search charge, while other models default to DeepSeek Search API. For teams building agents in the harness, a model choice can now also determine a search integration path.

What changed: DeepSeek Harness v0.2.2-rc.5 adds server-side search at no per-search charge for DeepSeek models, while other models default to the Search API.

What happened

DeepSeek published dsh-v0.2.2-rc.5 of DeepSeek Harness on GitHub on November 18, 2026. The release is marked as a pre-release. Its short note labels the change an improvement and states that DeepSeek models use DeepSeek server-side search. It further says that this search is not charged on a per-search basis.

The same note establishes a separate default for models outside the DeepSeek family: other models use DeepSeek Search API by default. In practical terms, the harness is not presenting one identical default search route for every model it can run. A workflow using a DeepSeek model follows the server-side-search route described in the release notes; a workflow using another model follows the Search API default described there.

That is a narrow release, but it is a material behavior change for an agent harness. Search can be part of an agent’s tool use when it needs current information or external references. With rc.5, changing the model can also change the default mechanism through which that tool use happens. DeepSeek’s announcement does not call the two paths equivalent. It does not provide a quality comparison, performance measurements, availability figures, or examples of how either path behaves in an agent task.

The pricing statement is similarly specific. DeepSeek says the server-side search used by DeepSeek models carries no per-search charge. The notes do not specify whether account, model, token, service, or other charges may still apply. They also do not state search quotas, rate limits, geographic availability, or eligibility requirements. The release notes do not say whether a user can override the defaults, how the harness identifies a DeepSeek model, or what configuration controls the choice of search route.

The GitHub page links a full changelog from dsh-v0.2.2-rc.4 to dsh-v0.2.2-rc.5 and identifies the release as immutable. Its stated operational change remains concise: DeepSeek models use server-side search without a charge per search, while other models default to DeepSeek Search API. Because this is a release candidate, teams should regard it as behavior to test rather than a final commitment for a stable version.

What the archive adds

New Runtime’s earlier coverage followed the Harness release sequence: v0.2.1-rc.3 on October 24, the v0.2.2 alpha on October 28, and v0.2.2-rc.2 on November 12. This is a new fact in that continuing story: rc.5 introduces model-dependent default search routing and a no-per-search-charge claim for DeepSeek models.

Trend signal

This supports Harness architecture outlives model choice in a qualified way. Search is being expressed at the harness layer, but the release shows that the tool path still changes with the selected model family. The update therefore does not demonstrate frictionless model substitution; it identifies one part of the workflow that teams must verify when they switch models.

Why now

The change arrives during the v0.2.2 pre-release sequence, after releases that added plugin, sub-agent, collaboration, and review-oriented features. It matters immediately to teams evaluating rc.5 because selecting another model can now change the harness’s default search route as well as the model itself.

Connections to other events

An earlier New Runtime note on November 12 covered v0.2.2-rc.2’s Git worktree review, named nested sub-agents, and session-based AgentTeam collaboration. Rc.5 adds a different operational layer: the external-search path available to those kinds of agent workflows. Together, the releases show the project expanding both orchestration features and the tools used inside a task.

Why it matters

For agent teams, search is not just a checkbox. It can affect tool permissions, request paths, costs, failure modes, and the reproducibility of a workflow. DeepSeek’s rc.5 note makes one of those choices explicit: DeepSeek models use server-side search, while other models default to DeepSeek Search API. The stated removal of a per-search charge may lower the barrier to experimenting with search-enabled workflows on DeepSeek models.

The broader lesson is not that one route is better. DeepSeek has not published evidence for that conclusion. Instead, builders should treat search as a model-dependent part of the harness configuration. A team that changes models for quality, latency, or cost reasons may also be changing the route that supplies external information. That can alter operational and financial assumptions even when the agent prompt and task remain unchanged.

What happens next

  • DeepSeek will clarify configuration, limits, or override behavior for the two search routes in a later release or documentation update (3 months, 55%). Watch for: release notes or configuration documentation naming search settings.
  • A stable v0.2.2 release will retain some form of model-dependent search routing (6 months, 60%). Watch for: stable release notes repeating or revising the rc.5 behavior.

Limits and uncertainty

DeepSeek’s release notes are brief and have not established how server-side search is configured or whether users can select a different route. The meaning of “no per-search charge” is limited to the stated billing unit; it does not prove unlimited use or the absence of other costs. No independent tests are reported for relevance, latency, reliability, coverage, or tool-call behavior. The announcement also does not disclose quotas, supported regions, account requirements, or whether the two routes return comparable results. As a pre-release, rc.5 may change before a stable version ships.

Skill to build now

Add search-route checks to a model-switch test plan this week. Run one representative agent task with a DeepSeek model and one with a non-DeepSeek model in the same Harness environment. Record which route is used, whether the task completes, the observed response time, and the billed components visible to your account. Use this only as a local validation exercise: the release notes do not promise identical output or cost behavior across the two paths.

Reader takeaway

DeepSeek Harness rc.5 turns search into a model-dependent default: DeepSeek models use server-side search with no per-search charge, while other models default to DeepSeek Search API. If a workflow relies on external information, validate the entire search path when changing models rather than treating search as an interchangeable background feature.

Open question

Will DeepSeek expose a user-controlled setting for search routing and publish limits or comparative behavior before v0.2.2 becomes stable?

Evidence / sources

  1. 1github.com/deepseek-ai/deepseek-harness/releases/tag/dsh-v0.2.2-rc.5
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicModelsExplore the models topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract