Blitz

POST /blitz

Auth: routing key (pllm_…). Note the path: blitz lives at /blitz on the base URL itself. There is no /v1/blitz.

Blitz runs a batch of prompts through the routing key’s lanes, under the same routing and budget rules as chat completions, and returns every result in one response. Built for bulk jobs like classification runs and evals. The gateway runs up to 8 prompts at a time and everything must finish within 5 minutes.

Request

curl https://api.proxyllm.ai/blitz \
  -H "Authorization: Bearer pllm_your_routing_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "flagship",
    "max_cost_usd": 2,
    "prompts": [
      { "messages": [{ "role": "user", "content": "Summarize: ..." }] },
      { "messages": [{ "role": "user", "content": "Classify: ..." }], "temperature": 0 }
    ]
  }'

Field	Type	Notes
prompts	array, required	1 to 50 prompts. Each is { messages, model?, temperature?, max_tokens? }.
model	string	Default model for prompts that don’t set their own. Same handling as chat completions.
temperature	number	Default for prompts that don’t set their own.
max_tokens	number	Default for prompts that don’t set their own.
max_cost_usd	number	Spend cap for the whole batch. Checked before each prompt starts, so in-flight prompts can overshoot it slightly.

Response

HTTP status is 200 even when individual prompts fail; check per-prompt status in results. results is ordered by prompt index.

{
  "job_id": "e7a2c410-...",
  "status": "partial",
  "total_latency_ms": 8412,
  "total_cost_usd": 0.0231,
  "success_count": 1,
  "error_count": 1,
  "results": [
    {
      "index": 0,
      "status": "success",
      "content": "Here's the summary...",
      "model": "gpt-5.5",
      "provider": "codex",
      "usage": {
        "prompt_tokens": 412,
        "completion_tokens": 128,
        "total_tokens": 540,
        "cached_tokens": 0,
        "cost_usd": 0
      },
      "latency_ms": 3120,
      "prompt_id": "1f6b...",
      "is_repeat": false,
      "finish_reason": "stop",
      "classifier_category": null
    },
    {
      "index": 1,
      "status": "error",
      "error": "cost_cap_reached",
      "model": "flagship"
    }
  ]
}

• status - completed when every prompt succeeded, failed when none did, partial otherwise.
• job_id - ties the batch together; each successful prompt also appears as its own request in usage.
• prompt_id and is_repeat - the gateway tracks repeated prompts; is_repeat is true when the account has sent these exact messages before.
• classifier_category - set when the key runs in classifier mode: the category the prompt was routed under.
• Per-prompt errors carry the failing lane’s message, or cost_cap_reached for prompts skipped after max_cost_usd was hit.

Error responses use the flat shape: { "error": "...", "code": "..." }.

Errors

Status	Code	When
400	(none)	prompts missing or empty, more than 50, or a prompt has no messages.
400	no_providers	The key has no lanes.
401	auth_required	No token sent.
401	invalid_api_key	Unknown or revoked routing key.
402	budget_exceeded	Monthly budget exhausted.
405	(none)	Method other than POST.
500	server_error	Creating the batch job failed.
500	unhandled	Unexpected crash.

Full list: error codes.