Gemini API with live pricing, in Google's own format

1. Create a key

   Sign up, add credit from $5, and create a key in API keys. Give keys that run unattended a spend limit.

2. Point the Gen AI SDK at Lovely Router

   Pass https://lovelyrouter.com as the SDK's base URL. The key goes in x-goog-api-key, which the SDKs send; ?key= and Authorization: Bearer work too.

3. Call generateContent

   generateContent, streamGenerateContent and countTokens behave as Google documents them.

   cURLPythonNode.js

   Shell

   curl "https://lovelyrouter.com/v1beta/models/gemini-3.1-pro:generateContent" \
     -H "x-goog-api-key: $LOVELY_API_KEY" \
     -H "content-type: application/json" \
     -d '{
       "contents": [{"role": "user", "parts": [{"text": "Explain DNS in two sentences."}]}]
     }'

Prices are in US dollars per million tokens, from the live catalog. Ids follow GET /v1/models, for example gemini-3.1-pro and gemini-3-flash, and Google's spellings such as gemini-3.1-pro-preview or models/gemini-3-flash resolve to the same model. GET /v1beta/models lists every model available in this format, in Gemini's own response shape. Gemini models are served in the gemini-value channel group; the model marketplace shows live availability and status hour-by-hour availability per channel group.

These Gemini endpoints are not served: cachedContents (explicit context caching), file uploads, embedContent and batchEmbedContents, batch mode and the Live API.

Gemini models also accept OpenAI Chat Completions and Anthropic Messages, which helps when your app only speaks one of those. thinkingConfig and other Gemini-only options are only certain to work in the native format.

The Gemini CLI reads its endpoint and key from the environment:

export GOOGLE_GEMINI_BASE_URL=https://lovelyrouter.com
export GEMINI_API_KEY=$LOVELY_API_KEY
gemini -m gemini-3.1-pro

Use with your tools covers other clients, and AI coding agents shows the Gemini CLI next to Claude Code and Codex CLI on the same key.

Gemini caches repeated prompt prefixes on its own, with nothing to mark. Cached tokens appear in usageMetadata.cachedContentTokenCount and are billed at the cache-read price, with no cache-write charge.

A cache only helps if the next turn reaches the channel that holds it. Lovely Router fingerprints each conversation from its system prompt and first user message and sends later turns back to the same channel. If your app rewrites or summarizes old history, send a stable x-session-id header per conversation instead. A timestamp in the system instruction breaks both the fingerprint and Gemini's own cache, so put it in the latest user message. Details are in Prompt caching.

Failures you would otherwise handle yourself are retried by the gateway first. A request that fails on a channel is retried on another one in the same group, up to three attempts, and a stream is held until its first token, so a failure before that point is retried instead of reaching you. What remains comes back in Gemini's error shape: RESOURCE_EXHAUSTED for 429 and UNAVAILABLE for 502. Retry those with backoff, and honor retry-after when it is present. Invalid requests (400, 404, 413, 422) are not retried, because they would fail the same way elsewhere.

Prefer alt=sse. Without it, streamGenerateContent returns one JSON array; with it, a stream that stalls for three minutes ends with a clean DEADLINE_EXCEEDED error event you can detect. Rate limits & failover has the full table.

Gemini is billed per token. Uncached input is promptTokenCount minus cachedContentTokenCount; cached tokens are billed at the cache-read price; output is candidatesTokenCount plus thoughtsTokenCount, at the output price. Every channel group bills at list price today, so each Gemini model costs what Google lists. Charges are deducted from your prepaid balance within seconds of the response.

There is no subscription and no monthly fee. Requests refused before reaching a model cost nothing, retried attempts are not billed, and a stream you cancel pays only for what it produced. Pricing shows every product on one page.