Zabio now speaks Anthropic: Claude Code and the Anthropic SDK on one key
Zabio now serves Anthropic's Messages API next to its OpenAI compatible API. Point Claude Code or the Anthropic SDK at api.zabio.ai and reach Claude, GPT, Gemini, DeepSeek, Qwen and the rest with one key and one bill.
The video loads from YouTube only when you press play.
Anyone who works in two languages knows the feeling. Half of the room is closed to you until you speak their language. Then you learn it, and the whole room opens up.
Until today, Zabio spoke one language: OpenAI’s. That covers a lot of code. The OpenAI SDKs, LangChain, LlamaIndex, Vercel’s AI SDK and most agent frameworks can talk to any OpenAI compatible endpoint, so they could talk to Zabio.
But a big part of the room speaks Anthropic. Claude Code. The Anthropic SDKs for Python and TypeScript. Every agent and internal tool that was written against Claude first. Those tools couldn’t use Zabio without a rewrite.
Starting today, they can. Zabio now serves Anthropic’s Messages API on the same host and the same key as the OpenAI compatible API. Your account, your credit and your bill stay the same. You just get to speak both languages.
What changes in your code: two settings
The Anthropic SDKs add /v1/messages on their own, so the base URL is the API host without /v1:
- Anthropic SDKs and Claude Code:
https://api.zabio.ai - OpenAI SDKs:
https://api.zabio.ai/v1
Then send your Zabio key the way the SDK already sends keys. That is the whole migration:
import anthropic
client = anthropic.Anthropic( base_url="https://api.zabio.ai", api_key="insert your zabio.ai key here",)
message = client.messages.create( model="claude-sonnet-5", max_tokens=1024, messages=[{"role": "user", "content": "Hello"}],)print(message.content[0].text)The answer is Anthropic’s message, or Anthropic’s stream of events with stream=True. Errors come back in Anthropic’s envelope, so the SDK raises its own exceptions and your error handling keeps working. POST /v1/messages/count_tokens counts tokens, and GET /v1/models answers in Anthropic’s shape when the SDK asks for it.
Claude Code on Zabio
Claude Code takes three environment variables:
export ANTHROPIC_BASE_URL="https://api.zabio.ai"export ANTHROPIC_AUTH_TOKEN="insert your zabio.ai key here"export ANTHROPIC_API_KEY=""Start claude and it runs on your Zabio balance. Use /model inside a session to switch to any model id Zabio serves, or set ANTHROPIC_MODEL=auto and let Zabio pick the model for every request. Claude Code’s background calls, the ones that name dated ids like claude-haiku-4-5-20251001, are answered by the same model without the date.
Name a Claude model and nothing gets lost
When you name a Claude model in model (claude-sonnet-5, claude-opus-5-5, claude-haiku-4-5), Zabio sends your request body to Anthropic as you wrote it: your system blocks, your cache_control markers, thinking blocks with their signatures, images and documents. Prompt caching works the way you set it up, and cache reads and writes are billed at Anthropic’s own cache rates, the one hour cache included.
What you get on top is everything a Zabio key already gives you: one balance for every vendor, spend limits per key and per member, logs of every call, and fallback to another Claude model when the one you named has trouble.
Or send auto, and let your Anthropic code use any model
This is the part I like the most. Send auto, auto/cheap, auto/balanced or auto/best, or name another vendor’s model like gpt-5-mini, gemini-3.5-flash or qwen3.6-plus, and Zabio translates the request, routes it like any other call and translates the answer back into Anthropic’s shape.
Your code doesn’t know the difference. It still reads message.content, still gets tool_use blocks for tool calls, still streams content_block_delta events. But the request was answered by whichever model fits the job at the best price, and that can be a Claude model too.
For a team that started with Claude, that is the cheapest way I know to stop sending every small job to a frontier model: change the base URL, change the model to auto, and watch the saving on your dashboard.
Images work on both paths. Send an image block to a Claude model and it goes as you sent it; send it to auto and it reaches the vision models Zabio picks from.
The small print
A few things are Anthropic’s own services, and Zabio doesn’t resell them:
- Tools that run on Anthropic’s side and are charged per use (web search, web fetch, code execution, Claude Code’s advisor) are removed from the request and listed in
zabio.ignored_parameters. The call goes on with your other tools. Use custom tools for what you need the model to call. - Code execution containers, the MCP connector, fast mode and data residency are not served. Run your MCP client yourself and send its tools as custom tools.
- On
autoand on other vendors’ models, thinking blocks are not returned andcache_controlmarkers are not sent, because other vendors can’t read Anthropic’s signatures and cache repeated prefixes on their own. Name a Claude model when you need them.
The full list, with every field, is on the Anthropic SDK and Claude Code page of the docs.
Two languages, one bill
Every call on the Anthropic path is billed exactly like the same call on chat completions: the vendor’s list price plus Zabio’s fee, and on auto, the small fee for automatic model choice. One key, one prepaid balance, one bill, whichever language your code speaks.
If you’ve been running Claude Code or an Anthropic based app and wondering what it would cost on the right model for each request, now you can find out by changing two settings. Create a free account, point your base URL at https://api.zabio.ai, and tell me how it goes at hello@zabio.ai.