From Anthropic

This one is not a one-line change

Eterial serves no Anthropic-compatible endpoint. There is no /v1/messages, and the anthropic SDK cannot be pointed at Eterial by changing a base URL. Moving over means rewriting the calls against the OpenAI protocol.

The good news is that the OpenAI shape is the one every other provider also speaks, so this is a rewrite you only do once.

Swap the SDK

# pip install openai
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ETERIAL_API_KEY"],
    base_url="https://chat.eterial.ai/v1",
)

Authentication changes with it: x-api-key and anthropic-version are replaced by a single Authorization: Bearer header.

The shape differences that actually bite

Anthropic	OpenAI / Eterial
`system` as a top-level field	a `{"role": "system"}` message, first in `messages`
`max_tokens` required	`max_tokens` optional
`content` as a list of blocks	`content` as a plain string, or a list of parts
`stop_reason: "end_turn"`	`finish_reason: "stop"`
`stop_reason: "max_tokens"`	`finish_reason: "length"`
`usage.input_tokens` / `output_tokens`	`usage.prompt_tokens` / `completion_tokens`
response `content[0].text`	`choices[0].message.content`

A request, before and after

message = client.messages.create(
    model="claude-...",
    max_tokens=1024,
    system="You are terse.",
    messages=[{"role": "user", "content": "Hi! How are you doing today?"}],
)

print(message.content[0].text)

Tool calling is the biggest rewrite

Anthropic defines a tool with input_schema and returns a tool_use content block, which you answer with a tool_result block inside a user message. The OpenAI protocol defines a tool with function.parameters, returns message.tool_calls, and expects a separate message with role: "tool" carrying tool_call_id.

The arguments also differ in kind: Anthropic gives you a parsed object, the OpenAI protocol gives you a JSON string you have to parse yourself. See Tool calling.

Streaming is a different event model

Anthropic emits typed events — message_start, content_block_delta, message_delta, message_stop. The OpenAI protocol emits one chunk shape carrying choices[].delta, terminated by [DONE]. If you wrote an event-type switch, it goes away. See Streaming.

Which model to reach for

There is no Claude here, and no drop-in equivalent to claim there is. Start with minimax-m2.7 for general work, kimi-k2.6 for simple coding tasks, and deepseek-v4-flash where the volume is high and the work is simple, then compare on your own traffic — the catalogue lists what is available and what each model can do.