Route AI Models in DeepSeek Harness: DeepSeek, OpenAI, Claude and Gemini
A practical DSH model-routing guide: configure providers, switch models safely, compare them under the same tools and workspace, and escalate expensive coding tasks only when needed.
On this page
DSH is not limited to one model provider. The official provider guide documents a built-in DeepSeek route, catalog providers such as Anthropic and OpenAI, native-auth providers such as Vertex, and custom providers for gateways or providers missing from the installed catalog.
First: what model routing means in DSH
The official UI gives you configured providers and a model picker. Selecting a model also makes it the default for new sessions.
A session that has already sent a request keeps the model recorded in its own log. For a clean model comparison, start a new session instead of assuming an old session silently changed models.
The official provider guide describes provider/model selection, not a built-in cost-versus-quality auto-router. The routing policy below is an operating pattern: choose or escalate models deliberately unless you build automation around it.
Official basis: DeepSeek Harness Provider Guide ↗
Configure the providers you actually need
DeepSeek
Use the built-in DeepSeek card in Settings → Models. Save the API key and select an available model from the installed DSH catalog.
Anthropic / OpenAI
Choose Add provider and select a catalog provider such as Anthropic or OpenAI. The installed catalog supplies its endpoint, protocol and model list; you provide the credential.
Vertex / Bedrock / Azure / Codex
These use provider-native authentication. The official guide calls out ADC/project for Vertex, AWS credentials/region for Bedrock, api-version for Azure and OAuth for Codex; an API-key field alone is not enough.
Gemini and other routes
Use the route available in your installed catalog or a supported native provider such as Vertex. If the provider or gateway is absent, use Add a custom provider instead of inventing a model ID under the wrong provider.
Company gateway / self-hosted endpoint
Use Add a custom provider. Give it a permanent lowercase Provider ID, base URL, protocol, credential and at least one model. Fetch available models only works when the endpoint exposes the compatible model-discovery route.
Sources: Provider Guide · llm-pi-ai multi-provider adapter
Switch models without corrupting your comparison
Use the same workspace
Keep repository state constant. If one model sees different files or a different commit, the comparison is already contaminated.
Use the same DSH preset and tools
Keep the same mode, tool surface, permissions and task instructions. A stronger harness can make a weaker model look better.
Start a fresh session per model
Because an existing session retains its recorded model, use a fresh session when moving from DeepSeek to OpenAI, Claude or another route.
Keep acceptance criteria identical
Use the same tests, lint command, expected files and review checklist. Judge accepted output, not how convincing the prose sounds.
Record cost and retries
Track model spend, tool loops, failed attempts and human review time. The cheapest token price is not always the cheapest accepted task.
A practical three-tier coding-agent policy
Do not send every task to the most expensive frontier model. Start with the cheapest model that is likely to pass your objective acceptance checks, then escalate on clear failure signals.
| Tier | Model strategy | Typical work |
|---|---|---|
| Tier 1 — cheap and verifiable | DeepSeek V4.1 Flash or another low-cost configured model | Repository reading, docs, tests, repetitive refactors, structured extraction, low-risk batch changes |
| Tier 2 — richer modality / stronger middle tier | A capable Flash-class or multimodal model | Screenshot/PDF-heavy workflows, multimodal inspection, tasks that failed Tier 1 but remain easy to verify |
| Tier 3 — frontier escalation | GPT-6 Astra, Claude Fable 5.1 or your strongest configured model | Complex architecture, difficult debugging, high-risk changes, repeated tool failures, expensive human review |
Escalate when the task tells you to
- Tests or acceptance checks fail after the allowed retry budget.
- The model repeats the same tool error or edit loop.
- The requested change touches security, migrations, production infrastructure or other high-review-cost surfaces.
- The model cannot form a stable plan across a large multi-file change.
- Human review cost is becoming larger than the savings from the cheaper model.
Without a retry budget and escalation rule, a cheap model can turn low token price into high total cost through repeated tool loops.
A small A/B test you can run in DSH
Pick a disposable repository task that has an objective result. Run the same prompt in separate fresh sessions, changing only the selected model or provider.
Task accepted?
Elapsed time
Model/API cost
Number of tool calls
Retries / failed tests
Minutes of human review
If Model A is 20× cheaper per token but needs three retries and 15 minutes of human repair, it may not be the cheaper production choice.
FAQ
Can DeepSeek Harness use OpenAI and Claude?
Yes. The official provider guide documents catalog providers such as OpenAI and Anthropic in Settings → Models, in addition to the built-in DeepSeek route.
Can DSH use Gemini?
Use the Google/Vertex route available to your installed catalog or another supported route. Vertex uses native ADC/project authentication. If the provider or gateway you need is absent from the installed catalog, DSH supports a custom provider.
Does changing the selected model change an old session?
Not after that session has already sent a request. Official docs say the session keeps the model recorded in its log; model selection becomes the default for new sessions.
Does DSH automatically choose the cheapest or strongest model?
The official model-selection guide describes configured providers and a model picker, not an automatic cost/quality router. Treat escalation as a manual operating policy unless you implement your own workflow or plugin automation.
What is the fairest way to compare models in DSH?
Use fresh sessions with the same workspace, preset, permissions, prompt and acceptance tests. Record accepted-task cost, not just token price or a single benchmark score.
Compare current model performance and prices
Open the model hub for the current DeepSeek, GPT, Claude and Gemini evidence snapshot before deciding which providers deserve a place in your routing stack.
DeepSeek Harness — Configure models · @deepseek-ai/dsh-llm-pi-ai