OpenRouter Auto Router vs Auto Code: Settings and Costs

OpenRouter Auto Router keeps the openrouter/auto slug after the November 6, 2026 announcement. The same announcement adds openrouter/auto-code, a coding policy on the same engine. Both bill at the selected model's standard rate; model selection is part of that price.

Choose the general slug for mixed traffic, apps and general agents. Choose the coding slug for coding harnesses and coding agents, then evaluate it on your own tasks before making it a default.

Auto Code's model page, checked November 6, 2026, gives a November 5, 2026 release date, one day before the announcement. The source material is OpenRouter's announcement and docs, all checked November 6, 2026; we have not run a paid OpenRouter test.

Auto Router vs Auto Code at a glance

The two slugs share an engine, the same plugin for allow and exclude lists, and the same billing rule. The shared plugin is auto-router; both read allowed_models and excluded_models from it.

The Auto Code model page says it is unavailable with the EU in-region option, and the Auto Router guide says the router runs only in the global data region (both checked November 6, 2026).

What changes for coding agents?

Auto Code is a second slug on the same engine, with a policy tuned for code. OpenRouter says it starts in a higher cost tier that includes stronger coding models, weighs speed more heavily and penalizes failures more harshly (Auto Code guide, checked November 6, 2026).

Its policy weighs measured pass rate, speed and cost per solved task, plus live prices and cache status. The general Auto Router uses a fast decision model to rate each request in one call, then picks a model and a reasoning effort (Auto Router guide, checked November 6, 2026).

Inside an agent session, Auto Code keeps the current model while work is going well. It tries more reasoning effort before switching, and it switches sooner when tests fail or the same error repeats. OpenRouter says a switch can cost extra while the new model rebuilds the cache, so this behavior keeps more of the session on a warm model.

If you are comparing a fixed coding model instead, the coding-agent model comparison covers evaluation on repository tasks.

API setup: model and matching plugin ID

Set the model parameter to the slug you want, then send the matching plugin entry. Both Auto and Auto Code read allowed_models and excluded_models from the auto-router plugin.

This complete cURL request sends the general slug with a low cost tier:

curl https://openrouter.ai/api/v1/responses \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openrouter/auto",
    "input": "Explain when prompt caching helps",
    "plugins": [
      { "id": "auto-router", "cost_tier": "low" }
    ]
  }'

The low tier favors lower estimated cost, medium balances cost and capability, and high favors more capable models when the benefit justifies the cost. The strings xhigh and max are still accepted and behave like high (Auto Router guide, checked November 6, 2026).

The Auto Code request body, shown without the cURL wrapper, uses Chat Completions and puts the allowlist on the same plugin id:

{
  "model": "openrouter/auto-code",
  "messages": [
    { "role": "user", "content": "Add a retry with backoff to fetchUser in api/users.ts" }
  ],
  "plugins": [
    {
      "id": "auto-router",
      "allowed_models": ["anthropic/*", "openai/*"],
      "excluded_models": ["~anthropic/claude-opus-latest"]
    }
  ]
}

Patterns are wildcard-based and case-sensitive. An excluded model is skipped even when it matches an allowed pattern. If your restrictions leave an empty eligible set, the request returns HTTP 404.

For a general tier on Auto Code, the Auto Code guide uses a separate jev-router plugin entry instead:

{
  "id": "jev-router",
  "cost_tier": "medium"
}

Two more settings matter. provider.max_price filters endpoints after the router picks a model; the model choice stays with the router. Add session_id in the body, up to 256 characters, or send x-session-id in the header; the body value wins when both are present.

If your client expects a fixed model ID instead of a router slug, Cline's custom API key setup shows where it goes.

What each request costs

Model selection adds nothing to the bill: you pay the selected model's standard rate, and the router fee is zero (Auto Router and Auto Code guides, checked November 6, 2026). Auto Code is also priced by the model it selects.

The default general policy can cost more per task than the previous Auto Router. OpenRouter's guide recommends cost_tier: low to stay closer to the old behavior (checked November 6, 2026).

BYOK requests follow OpenRouter's separate billing policy; check the BYOK docs before mixing provider keys with either router (checked November 6, 2026). The router's selection fee is zero.

Caching, model changes and response metadata

Switching models mid-session costs extra while the new model rebuilds the input token cache (announcement, checked November 6, 2026). The docs say the router favors a change in reasoning effort when that is enough.

Sessions are sticky on a best-effort basis and expire after 60 minutes without a request (prompt caching guide). A successful request resets the timer.

For billing attribution, log the response model field: it names the model that actually served the request. Within openrouter_metadata, the top-level requested member shows what you asked for, which can differ; is_byok says whether your provider key was used.

Router metadata is opt-in: send X-OpenRouter-Metadata: enabled (metadata schema). Cache hits omit openrouter_metadata, so a cached replay shows the served model but omits the original selection details. The response also carries cache_discount and cached_tokens, though a cache hit remains best-effort.

How to read OpenRouter's benchmark claims

OpenRouter's announcement reports a vendor benchmark comparison: Auto Code scores in line with Claude Opus 5.5 across five benchmarks while costing 74% less per task (OpenRouter, November 6, 2026). The comparator is Auto Code against Opus 5.5; the general Auto Router is a different slug. Treat this as OpenRouter's per-task benchmark claim; compare your own tasks before budgeting.

Choose automatic selection or a fixed model

Use the general Auto Router when request types vary and you want one slug across an app, chat product or general agent. Use Auto Code in a coding harness or coding agent, the kind of work its policy is designed for, and evaluate it on your own tasks before making it a default.

Choose a fixed model when you already know the model that works for your work. OpenRouter's announcement recommends that, and a fixed ID also makes cost, caching and behavior easier to compare across releases. For the tradeoff between automatic selection and a fixed model, see the OpenRouter alternative comparison.

OmniaKey's model catalog lists fixed models available through one key.

For an agent, evaluate the fixed model on your own repository tasks rather than on a leaderboard.

Frequently asked questions

Is Auto Code a different model from the Auto Router?

No. It is a second slug on the same engine, with a policy tuned for coding work, and it selects from the models you allow. The slug is openrouter/auto-code.

Does Auto Code respect cost_tier on the auto-router plugin?

No. It ignores cost_tier there and runs its own coding tier. To ask for a general low, medium or high tier instead, set cost_tier on a separate jev-router plugin entry.

Which one costs less per request?

Both are priced at the serving model's standard rate, and the selection fee is zero. Which costs less depends on the model each one selects for your tasks, so compare them on your own tasks.

How do I see which model actually answered?

Read the response model field; it names the serving model. For per-request detail, opt in with the metadata header.

Can I limit Auto Code to specific models?

Yes. Put allowed_models and excluded_models on the auto-router plugin; matching is case-sensitive and exclusions win. A list that leaves an empty eligible set returns HTTP 404.

Sources checked

• OpenRouter: new Auto Router and Auto Code announcement (November 6, 2026)
• OpenRouter: Auto Router guide
• OpenRouter: Auto Code guide
• OpenRouter: Auto Code model page
• OpenRouter: router metadata
• OpenRouter: prompt caching and sticky sessions
• OpenRouter: BYOK pricing

All checked on November 6, 2026.

Update notes

• November 6, 2026 source snapshot: OpenRouter's announcement and docs checked that day. The Auto Code model page gives a November 5, 2026 release date; the announcement carries a November 6, 2026 byline.
• Recheck the plugin ids, cost tiers and Auto Code policy whenever OpenRouter updates its changelog or router docs.