Native protocols

This page is for developers who already use the Anthropic Messages or Google Gemini protocol and want to keep that request shape.

Anthropic Messages#

Use the base URL https://api.clssai.com without /v1, then call POST /v1/messages; POST /anthropic/v1/messages is an equivalent path. Authenticate with x-api-key or Bearer and always send the required anthropic-version: 2023-06-01 header. The gateway supplies metadata.user_id only when missing and removes a beta query parameter. Send optional beta declarations through the standard Anthropic header instead.

Anthropic routes use Anthropic-native model names such as the verified claude-haiku-4-5. The gateway also maps a few dotted aliases (claude-opus-4.8 resolves to claude-opus-4-8), but coverage is partial: claude-haiku-4.5 returns 404, and OpenAI-compatible IDs such as anthropic/claude-haiku-4.5 are never valid here. Send the hyphenated native name from public GET /anthropic/v1/models and confirm it in the response model field.

curl https://api.clssai.com/v1/messages \
  -H "x-api-key: $CLSSAI_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-haiku-4-5","max_tokens":128,"messages":[{"role":"user","content":"Reply with OK."}]}'

The Anthropic SDK appends /v1/messages, so its base_url must not contain /v1.

import os
from anthropic import Anthropic

client = Anthropic(
    base_url="https://api.clssai.com",
    api_key=os.environ["CLSSAI_API_KEY"],
)
message = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=128,
    messages=[{"role": "user", "content": "Reply with OK."}],
)
print(message.content[0].text)

Prompt caching#

The gateway passes cache_control through unchanged and does not rewrite it. In a verified claude-haiku-4-5 request, the first response reported cache_creation_input_tokens: 4812 and cache_read_input_tokens: 0. The next three consecutive requests kept the same system prefix and changed only the user message; all three reported cache_read_input_tokens: 4812.

Upstream models impose a minimum cacheable prefix length. The Haiku tier has a 2048-token minimum; a shorter prefix does not create a cache entry. This threshold is set by the upstream model, not by the gateway.

Count tokens#

POST /v1/messages/count_tokens uses a dedicated upstream credential. These calls are not billed and are not written to request logs. The anthropic-version header is required here too.

curl https://api.clssai.com/v1/messages/count_tokens \
  -H "x-api-key: $CLSSAI_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-haiku-4-5","messages":[{"role":"user","content":"Count this input."}]}'

Google Gemini#

Use the base URL https://api.clssai.com/v1beta. The gateway forwards Gemini request bodies without rewriting them. Send the key with the standard x-goog-api-key header or a ?key= query parameter. Bearer authentication is not accepted. Gemini routes use Google-native model names; gemini-3.1-flash-lite has been verified. List current names with public GET /v1beta/models.

curl https://api.clssai.com/v1beta/models/gemini-3.1-flash-lite:generateContent \
  -H "x-goog-api-key: $CLSSAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Reply with OK."}]}]}'

import os
import requests

response = requests.post(
    "https://api.clssai.com/v1beta/models/gemini-3.1-flash-lite:generateContent",
    headers={"x-goog-api-key": os.environ["CLSSAI_API_KEY"], "Content-Type": "application/json"},
    json={"contents": [{"parts": [{"text": "Reply with OK."}]}]},
    timeout=120,
)
response.raise_for_status()
print(response.json())

const response = await fetch(
  "https://api.clssai.com/v1beta/models/gemini-3.1-flash-lite:generateContent",
  {
    method: "POST",
    headers: { "x-goog-api-key": process.env.CLSSAI_API_KEY, "Content-Type": "application/json" },
    body: JSON.stringify({ contents: [{ parts: [{ text: "Reply with OK." }] }] }),
  },
);
if (!response.ok) throw new Error(await response.text());
console.log(await response.json());

Common mistakes#

• Do not add /v1 to the Anthropic SDK base_url; the SDK adds it.
• Do not omit the required anthropic-version: 2023-06-01 header.
• Do include /v1 in an OpenAI SDK base URL.
• Authenticate Gemini with x-goog-api-key or ?key=, not a Bearer token.
• Use GET /anthropic/v1/models or GET /v1beta/models and send the protocol's native model name. Do not reuse an OpenAI-compatible author/model ID.
• Run SDK and fetch calls on a server; a browser call exposes the key to every visitor.
• Check empty OpenAI-compatible streaming choices arrays and use cf-ray to trace inference.