BlockRun

Rate Limits

BlockRun is a pass-through gateway: for paid model calls it does not add its own per-request throttle. The rate limits you may hit are the upstream provider's capacity limits (tokens-per-minute / requests-per-minute on the provider tier backing that model). When an upstream provider throttles a request, BlockRun surfaces it to you transparently as an HTTP 429 so you can back off or fail over.

The 429 response

When an upstream provider rate-limits a request, BlockRun returns:

HTTP/1.1 429 Too Many Requests
Retry-After: 60
X-RateLimit-Source: anthropic
Content-Type: application/json

{
  "error": "Rate limited",
  "message": "Upstream provider rate limit hit — retry after 60s, or fail over to a same-tier model on a different provider.",
  "code": "RATE_LIMITED",
  "source": "anthropic",
  "retry_after_seconds": 60
}
Field / HeaderMeaning
Retry-After (header)Seconds to wait before retrying. Honor this.
X-RateLimit-Source (header)Which upstream provider throttled (anthropic, openai, …).
codeAlways RATE_LIMITED for this case.
retry_after_secondsSame value as Retry-After, in the body for convenience.

This applies to both the standard (POST /api/v1/chat/completions) and Anthropic-compatible (POST /api/v1/messages) endpoints, for streaming and non-streaming requests. For streaming, the 429 is returned before the first SSE byte (no partial stream is emitted).

Recommended client handling

  1. Honor Retry-After — wait the indicated seconds, then retry (exponential backoff on repeats).
  2. Or fail over to a same-tier model on a different provider — e.g. if anthropic/claude-sonnet-4.6 is throttled, retry on openai/gpt-5.4 or google/gemini-3-pro-preview. Different providers have independent rate-limit pools, so a cross-provider retry usually succeeds immediately.
import time
resp = client.chat(...)
if resp.status_code == 429:
    time.sleep(int(resp.headers.get("Retry-After", 60)))
    resp = client.chat(...)            # retry
    # or: client.chat(model="openai/gpt-5.4", ...)  # cross-provider failover

Provider notes

  • Claude (anthropic/*) is served through AWS Bedrock (cross-region inference) with a fallback to the direct Anthropic API. A 429 here means both pools were saturated; back off or fail over to another provider.
  • GPT (openai/*) mainline chat models are served Azure-first with a fallback to direct OpenAI.

In both cases the failover is automatic and internal — you only see a 429 when every backing pool for that model is exhausted.

Other endpoints

Some non-LLM endpoints (image generation, async job submission, wallet reconciliation, RealFace init) carry small per-IP limits to bound abuse and real upstream cost. When exceeded they also return 429; the same Retry-After guidance applies.

Links