Claude SDK

import anthropic

client = anthropic.Anthropic(
    base_url="https://api.vectoraxis.ai",
    api_key="vk-c7d4e1f2a3b95a8b.Rq8mT2vX5nL9pW3kY7cJ1hF4dS6gB0zN2aE8uV5tQ1c",  # an Anthropic virtual key
)

message = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
print(message.content[0].text)

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic({
  baseURL: 'https://api.vectoraxis.ai',
  apiKey: process.env.VECTORAXIS_KEY,   // vk-...
});

const message = await client.messages.create({
  model: 'claude-haiku-4-5',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Hello, Claude' }],
});

curl https://api.vectoraxis.ai/v1/messages \
  -H "x-api-key: vk-c7d4e1f2a3b95a8b.Rq8mT2vX5nL9pW3kY7cJ1hF4dS6gB0zN2aE8uV5tQ1c" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-haiku-4-5","max_tokens":1024,"messages":[{"role":"user","content":"Hello, Claude"}]}'

Every VectorAxis feature is a request header, exactly as on the OpenAI-compatible endpoint. Set them once with the SDK’s default_headers (Python) / defaultHeaders (TypeScript), or per call with extra_headers. See Request headers.

The Python and TypeScript SDKs send their API key as x-api-key; the Java SDK sends it as Authorization: Bearer. Both are accepted. In order, the gateway reads:

A virtual key for another provider works too: the gateway translates the Messages request to that provider’s format and the answer back, so an application written for the Claude SDK can run on an OpenAI model by changing only its key and model.

Both endpoints hand the gateway the same request, so everything a virtual key enforces applies here:

• Key checks: status, expiry, credit limit, model allowlist, and requests and tokens per minute and day.
• Plan limits: the workspace request rate.
• Guardrails: input and output, the key’s own plus any named in x-guardrail. A block is a 422 with x-guardrail-status: BLOCK, reported as invalid_request_error.
• Tool-result scan: at the key’s own action, with x-tool-scan-status: BLOCK on a block.
• Gateway features: response caching, retries and fallbacks, routing configs, prompt attribution (x-prompt-slug), logs and cost.

Errors always come back in Anthropic’s shape, {"type":"error","error":{"type":…,"message":…}}, so the SDK raises its usual exception classes. A provider error keeps the provider’s status, type and message.

Every cache_control breakpoint reaches Anthropic on the block it was on, and the response carries Anthropic’s own usage split: cache_creation_input_tokens, cache_read_input_tokens and cache_creation by TTL. This is Anthropic’s prompt cache, separate from the gateway’s response cache: a hit still calls the model, but the cached prefix is billed at the cache rate.

message = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=1024,
    system=[{
        "type": "text",
        "text": LONG_REFERENCE_TEXT,
        "cache_control": {"type": "ephemeral"},          # or {"type": "ephemeral", "ttl": "1h"}
    }],
    messages=[{"role": "user", "content": "Summarise section 3."}],
)
print(message.usage.cache_creation_input_tokens, message.usage.cache_read_input_tokens)

Cost: cache reads, cache writes and 1-hour writes are charged at the model’s own cache rates (on Claude, a read is about a tenth of the input rate and a 5-minute write 1.25 times it), the rest at the input rate. A model must see a minimum prefix before it caches anything (4,096 tokens on Haiku 4.5); shorter prefixes are not cached and report zero cache tokens.

POST /v1/messages/count_tokens returns {"input_tokens": N} from Anthropic’s own counter, for Anthropic keys and Claude on Vertex AI. It takes the same credentials and key checks as /v1/messages. Nothing is generated, so nothing is billed or logged. Another provider, or a routing config, is a 400.

count = client.messages.count_tokens(
    model="claude-haiku-4-5",
    system="You are a weather assistant.",
    tools=[weather_tool],
    messages=[{"role": "user", "content": "What is the temperature in Lisbon?"}],
)
print(count.input_tokens)

The Batches API runs many Messages requests asynchronously at half the standard price, usually within minutes and always within 24 hours. Create, list, retrieve, cancel, delete and results all work through the SDK’s messages.batches.

batch = client.messages.batches.create(requests=[
    {"custom_id": "ticket-1", "params": {"model": "claude-haiku-4-5", "max_tokens": 256,
        "messages": [{"role": "user", "content": "Classify: 'I was charged twice.'"}]}},
    {"custom_id": "ticket-2", "params": {"model": "claude-haiku-4-5", "max_tokens": 256,
        "messages": [{"role": "user", "content": "Classify: 'How do I reset my password?'"}]}},
])

# later, once batch.processing_status == "ended"
for result in client.messages.batches.results(batch.id):
    print(result.custom_id, result.result.type)

These are Anthropic’s batches. OpenAI’s file-based batches are /v1/batches, a separate API.

The SDK’s anthropic-beta header (its beta client, and Claude Code) is passed to Anthropic and Claude on Vertex AI on /v1/messages, count_tokens and /v1/chat/completions, retries and fallbacks included. It is never sent to another provider. A value that is not a comma-separated list of beta names (lower-case letters, digits and dashes, at most 32) is a 400 from the gateway; a well-formed name Anthropic does not know is Anthropic’s own 400.