Skip to content
Tapotik AI

Command Palette

Search for a command to run...

Rate limits

Limits are applied per API key across three dimensions: requests per minute (RPM), tokens per minute (TPM) and concurrent jobs. Your tier follows your plan and upgrades take effect immediately.

Limits by plan

PlanRequests / minTokens / minConcurrent jobs
Starter6050,0002
Pro600500,00020
EnterpriseCustomCustomCustom — dedicated clusters

"Jobs" are async workloads: image batches, long voice renders, video generations and agent runs. Chat completions count against RPM/TPM only. Enterprise plans run on dedicated inference clusters, so limits are negotiated to your traffic profile — see pricing.

Rate limit headers

Every response includes live quota information so clients can self-regulate before hitting a limit.

HeaderDescription
X-RateLimit-Limit-RequestsYour RPM ceiling for this key.
X-RateLimit-Remaining-RequestsRequests left in the current window.
X-RateLimit-Limit-TokensYour TPM ceiling for this key.
X-RateLimit-Remaining-TokensTokens left in the current window.
X-RateLimit-ResetSeconds until the window resets.
Retry-AfterOn 429 responses only — seconds to wait before retrying.

Handling 429s

When a limit is exceeded the API returns 429 rate_limit_exceeded. Respect Retry-After and retry with exponential backoff and jitter:

lib/backoff.tstypescript
async function withBackoff<T>(fn: () => Promise<T>, maxRetries = 5): Promise<T> {
  for (let attempt = 0; ; attempt++) {
    try {
      return await fn();
    } catch (err) {
      if (!(err instanceof Tapotik.RateLimitError) || attempt >= maxRetries) throw err;

      const retryAfter = Number(err.headers?.["retry-after"] ?? 0);
      const backoff = Math.min(2 ** attempt * 500, 30_000);
      const jitter = Math.random() * 250;

      await new Promise((r) => setTimeout(r, Math.max(retryAfter * 1000, backoff) + jitter));
    }
  }
}

const completion = await withBackoff(() =>
  client.chat.completions.create({ model: "tapotik-2-flash", messages })
);
The official SDKs retry 429 and 503 responses automatically (twice by default). Tune with maxRetries in the client constructor before rolling your own.

Staying under the limits

  • Stream long completions — streaming counts tokens as they generate, smoothing TPM spikes.
  • Use tapotik-2-flash or router/auto for high-volume traffic; flash tokens are cheaper and faster to schedule.
  • Batch image variations into a single n > 1 request instead of parallel single requests.
  • Split traffic across keys per service — limits are per key, and per-key budgets isolate noisy neighbors.
Limits protect shared capacity — they are not a throughput guarantee. If you consistently run above 80% of Pro limits, talk to us about Enterprise dedicated clusters before launch week, not during it.