Messages

The Anthropic-style Messages endpoint, parameters, response shape, and token counting

POST /v1/messages is the Anthropic-style endpoint. It is a drop-in for the anthropic SDK and Claude Code.

Request

Authenticate with x-api-key: sk-ablatic-.... The base URL for this surface is https://api.ablatic.ai.

cURL

curl https://api.ablatic.ai/v1/messages \
  -H "x-api-key: $ABLATIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "talos",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

Parameters

Name	Type	Required	Description
`model`	string	yes	Model id, nonempty
`messages`	array	yes	Conversation messages, 1 to 200
`max_tokens`	integer	yes	Max output tokens, 1 to 200000
`system`	string or array	no	System prompt
`metadata`	object	no	Request metadata
`stop_sequences`	array	no	Accepted but ignored
`stream`	boolean	no	Stream over the Anthropic SSE event graph
`temperature`	number	no	Sampling temperature, 0 to 2
`top_p`	number	no	Nucleus sampling cutoff
`top_k`	integer	no	Top-k sampling cutoff
`tools`	array	no	Tool definitions. Uses `input_schema`
`tool_choice`	object	no	Tool selection policy
`thinking`	object	no	`{ "type": ..., "budget_tokens": ... }`

The model name selects the reasoning mode by substring. A name containing haiku maps to fast, sonnet and opus to thinking (opus with the largest reasoning-token budget). Send talos for the default thinking mode. Unknown names map to thinking.

Response

{
  "id": "msg_...",
  "type": "message",
  "role": "assistant",
  "model": "talos",
  "content": [
    { "type": "text", "text": "Hello. How can I help?" }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 9,
    "output_tokens": 7,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0
  }
}

The model field echoes the model you sent. usage reports input_tokens, output_tokens, cache_creation_input_tokens, and cache_read_input_tokens.

Behavior notes

response_format is not honored on this surface. There are no structured outputs here. For JSON mode use Chat completions.

Prompt caching is real and billed. The server caches prompt prefixes automatically. On this surface cache_control marks which prefixes to prime and shapes the usage breakdown to match Anthropic. Cached reads are billed at a discount and the first write at the normal input rate, with no premium. See Usage and billing. Up to 4 cache_control breakpoints are kept, extras are truncated.
Prior thinking blocks in the input are dropped.
PDF document blocks are extracted. URL documents are not fetched.

Count tokens

POST /v1/messages/count_tokens returns an estimate.

{ "input_tokens": 42 }

The count is a heuristic (roughly characters divided by 4), not an exact token count.

Batches

POST /v1/messages/batches and its companion endpoints run Anthropic Message Batches. The body is requests[] of { custom_id, params }, with a 29-day result window, a cap of 100000 requests or 256 MiB per batch, and 50 in flight per user. They support Idempotency-Key, and results stream as application/x-ndjson once processing_status is ended. See Batch.