Agent usage and rate limits

Understand the included Chat/Agent budget, approximate request coverage, and request limits for chat-agent access.

Updated

Agent requests are included in paid Chat plans starting with Pro, but they are not unlimited. They consume the same included subscription budget as normal LLMBase Chat requests. Heavy agent usage therefore leaves less included budget for the same subscription period. If the included budget is exhausted, the API returns a budget error before running the model request.

Agent requests and normal LLMBase Chat requests draw from the same included Chat/Agent budget. The dashboard shows budget used and left as percentages for the current billing period; it does not expose the internal budget amount.

Approx. short text requests per month

The included budget is not a fixed request quota. For short text chats, the monthly budget usually covers approximately:

Model classFreeStarterProExpertMax
Small models500-1,5003,500-10,0007,500-20,00037,500-100,00075,000-200,000
Medium modelsNot included700-2,0001,500-4,0007,500-20,00015,000-40,000
Large modelsNot includedNot included150-800750-4,0001,500-8,000

Estimates are for short text chats. They are not fixed request quotas; long context, tools, images, research, and high-output answers use more of the included budget. Agent keys require Pro, so the Free and Starter columns are shown for browser Chat planning.

Budget reservations

Before running a Chat or Agent API request, LLMBase may temporarily reserve part of the included budget for that request. The reservation is based on the model, context size, enabled tools, and possible output length. This prevents multiple parallel requests from spending the same remaining budget at the same time.

Successful requests are finalized with actual usage. Failed or interrupted requests are released automatically. Very large requests can be rejected before the model runs if they do not fit the remaining included budget; shorten the context, reduce the requested output length, or choose a cheaper model.

Request limits

Request limits are separate from the included subscription budget. During high request volume or repeated automation, the API can return 429 with rate_limit_exceeded. Wait one minute before retrying, avoid parallel retry loops, and use exponential backoff for unattended clients. If an ordinary workflow is consistently rate-limited, contact support with a sanitized timestamp and request ID.

Request body size

POST /chat/completions on the agent API accepts a JSON body of up to 8 MiB (8,388,608 bytes). A larger body is rejected with HTTP 413. The error code stays request_body_too_large, and error.type is request_too_large. The error message names the 8,388,608-byte limit and tells the client to compact the session (/compact) or start a new one.

{
  "error": {
    "message": "Request body exceeds the 8388608-byte (8 MiB) limit. Compact the session (/compact) or start a new one.",
    "type": "request_too_large",
    "code": "request_body_too_large"
  }
}

Clients that understand request_too_large, including OpenCode V2, can treat that 413 as a context overflow, compact, and retry once. A 413 without that type does not start OpenCode V2’s automatic recovery. See the OpenCode example for limit.context and compaction.auto.

This cap is separate from the included subscription budget and from 429 rate_limit_exceeded. A request under 8 MiB can still be rejected when it does not fit the remaining included budget.

If you see request_body_too_large, shorten the conversation before retrying. Compact history (/compact) or start a new session for a long task, and send excerpts instead of whole logs, file dumps, or repeated instructions.

Production workloads

For deterministic production jobs, background workers, or usage that must scale with exact cost accounting, use the direct Inference API instead of the Pro agent subscription.

Agent API responses use the same OpenAI-compatible behavior, usage fields, and prompt-cache support as direct inference. LLMBase automatically adds a stable prompt_cache_key for subscription-agent calls when the client does not provide one.

Clients may still send their own prompt_cache_key if they need a stable cache namespace per workspace, repository, or long-running agent thread.

Starter and Free users still cannot use chat-agent keys, even if they have prepaid inference credits. For production jobs, unattended background workers, or customer-facing API products, use the direct Inference API with a llmbase_... inference key.