Start from what is missing, not from a ranking

Part 6 made the case that one configured route can be enough, especially when that route is OpenRouter. So the question here is not which company wins. It is: what is my current route missing?

Compare native API features, model availability, account controls and route independence. A universal ranking would be stale the week it was published, and it still would not tell you whether your current route covers your work.

flowchart TD
  A["OpenRouter already working"] --> B{"Can you name what is missing?"}
  B -->|"nothing"| C["Stay on one route"]
  B -->|"a native API feature"| D["A direct provider that documents it"]
  B -->|"separate billing or account controls"| E["A direct provider account"]
  B -->|"an independent path when a route fails"| F["A second route, different company"]
  D --> G["Match API mode, base URL,
thinking format and model ID"]
  E --> G
  F --> G
  G --> H["Test, then verify with a low-risk prompt"]

Three fields decide whether a route works at all

A provider entry is mostly ordinary: a name you choose, a key you paste. Three fields are not preferences — they are protocol settings, and no default covers every route.

• API mode — a supported protocol identifier. Most routes use openai-completions. Anthropic direct uses anthropic-messages. Other supported modes include openai-responses, google-generative-ai, mistral-conversations and bedrock-converse-stream.
• baseUrl — the endpoint for that route. An OpenAI-compatible protocol does not imply the OpenAI endpoint. That is the commonest wrong assumption when copying a configuration from elsewhere.
• thinkingFormat — the parser for route-specific reasoning data. Legal values include openai, openrouter, deepseek, together, baseten, zai, qwen, chat-template, qwen-chat-template, string-thinking and ant-ling. Anthropic direct uses native blocks through anthropic-messages and leaves this field blank.

Where you type them depends on which group you pick in the Add provider catalog.

Where the fields liveThe catalog opens on the Models tab of Settings: a search box, then every route sorted into three groups.

• CUSTOM is one card: OpenAI / Anthropic compatible, subtitled Custom endpoint format. This is the entry that asks you for the three fields above — so it is where all three can go wrong.
• SUBSCRIPTIONS holds six cards — Anthropic (Claude Pro/Max), GitHub Copilot, Kimi Code, OpenRouter OAuth, Radius and xAI — each marked OAuth. You connect by logging in; there is no endpoint to mistype.
• API KEY is the third group, and the heading is the authentication method: a key rather than a login. Each card carries a model count — Amazon Bedrock 121 models, Ant Ling 3, Anthropic 14 — which tells you how large a route is before you open it.

The fourth field is the model ID, and it is route-specific

The same model family answers to a different string depending on whether you reach it through a gateway or directly. Claude is the clearest example: all four fields differ between the two routes.

flowchart LR
  A["You want a Claude model"] --> B["Route A · OpenRouter gateway"]
  A --> C["Route B · Anthropic direct"]
  B --> D["mode: openai-completions
url: https://openrouter.ai/api/v1
thinking: openrouter
id: anthropic/claude-sonnet-4.5"]
  C --> E["mode: anthropic-messages
url: https://api.anthropic.com/v1
thinking: blank, native blocks
id: claude-sonnet-4-6"]

So copy the ID from the route you configured, never from an article or a marketing page — and watch the punctuation. Direct Anthropic naming can use 4-6 where an OpenRouter catalog entry uses 4.5, and a dated ID such as claude-haiku-4-5-20251001 pins a version on purpose.

All seven routes on one screen

This is the part worth coming back to. Protocol values are from the add-on baseline; model IDs are examples documented at the time of writing. Check each one against the provider's live catalog before you save — hosted models get retired and renamed without notice.

Six steps, without disturbing the route that works

This adds DeepSeek as a second route. Any other provider follows the same shape with its own documented API mode and endpoint — that is what the table above is for.

Two controls, two jobs: + Add Provider creates the entry, + Add Model adds an ID to one that already exists. Your OpenRouter entry is not edited or replaced.

1. Step 1

   Create a provider key

   Open https://platform.deepseek.com/ and follow the current account process. Create a key named pi-agent-home and store it in a password manager. Account requirements vary by provider and region.

2. Step 2

   Open the Models panel and choose Add Provider

   Keep the existing OpenRouter card. In the Models panel, select + Add Provider, or the equivalent control in your version.

3. Step 3

   Enter the four fields

   Name it DeepSeek, choose openai-completions, enter https://api.deepseek.com/v1, and paste the key. Use deepseek as the thinking format for the reasoning route.

4. Step 4

   Add the exact model IDs

   Select + Add Model, add deepseek-chat, then add deepseek-reasoner. Recheck both against the official model page before saving.

5. Step 5

   Run Test, then Save

   Select Test and wait for the result. If it fails, read the actual error before changing fields — guessing at the four values in turn is how a working route gets broken. Then select Save.

6. Step 6

   Verify with a low-risk prompt

   Select the new model in the composer and send something short. Confirm the expected route was used, then check the provider's own usage record.

Step 6 is the one people miss, because it happens outside Settings.

Where step 6 happensThe model list opens from the button under the message box. The check mark is on GPT-5.6 Sol.

• The check mark shows which model the next message will use — here, GPT-5.6 Sol. Adding a provider does not move it; you still have to come here and choose.
• The list is flat. Eight names on this instance, one line each. A second route just makes it longer — look for the models you added in step 4.
• It opens from the composer, not from the panel where you typed the four fields. Settings defines a route; the composer is where one gets used.

Pick by the job in front of you

Use this after checking the live catalogs, not instead of it.

Notice how often the first column is the route you already have. The catalog shows why.

Why the gateway staysOpenRouter, 366 models behind one key. Vercel AI Gateway, 237. Every other card on this screen is smaller than both.

• Radius shows 0 models. A card in the catalog is not a promise of models behind it — which is why step 4 has you paste exact IDs and step 5 has you test.

Where to go from here