Pricing, Credits & Rate Limits

What you pay and how much you can use, on one page: plans, credits, per-API prices, and per-API rate limits. All prices are in Indian Rupees (₹) and exclude GST. Usage is rounded as described in the billing notes for each service.

Plans

Choose or change a plan from the Sarvam Dashboard. For Enterprise pricing, contact us.

Credits

Sarvam offers ₹100 worth of free credits for every verified user on signup. These credits can be used across any of our APIs, explore, prototype, and build without upfront cost.

Credits are universal and never expire. Once exhausted, add more credits or upgrade your plan from the Sarvam Dashboard.

For commercial production rights when you ship generated audio (Bulbul v3, Content Studio exports, cloned voices), see Commercial Licensing.

If your credits are exhausted, API requests fail with 402 (insufficient_quota_error), which is separate from a 429 rate limit. You can add credits at any time, adding credits does not change your plan or rate limits.

1. Add Credits: Top up from the Billing page at any time. Credits never expire.
2. Upgrade Your Plan: Higher plans include bonus credits and increased rate limits.
3. Enterprise: For volume discounts and custom billing arrangements, email support@sarvam.ai.

Prices by API

Rates are listed for each self-serve plan.

Language Models

The open-weight models (Gemma, GLM, DeepSeek, Kimi, MiMo) are served by Token Factory.

All prices are ₹ per 1M tokens and are the same on every plan. Reasoning tokens, where applicable, are billed as output tokens.

Sarvam Models

Sarvam Prism is billed on input tokens only: output is ₹0 because it returns probabilities, not text. The state is counted once per request, so asking 3 short questions about a 1,000-token state bills a little over 1,000 input tokens. See Token usage.

Open-Weight Models

Speech to Text

Priced per minute, billed per second.

Text to Speech

Priced per 1,000 characters, rounded to the nearest character. Rates apply to REST, HTTP streaming, and WebSocket streaming.

Text Processing

Priced per 1,000 characters, rounded to the nearest character.

Document Intelligence

Priced per page.

Document Translation

Priced per 1,000 billable characters, per target language. Character count is computed after parsing; each target language in a job is billed separately. Charges round up to the nearest rupee; minimum ₹1 when any billable text is present.

Dubbing

Priced per minute of source media, multiplied by the number of target languages. Duration is rounded to whole seconds before billing.

Rate Limits

Rate limits restrict the number of API requests your account can make within a given time window. Key points:

• Per-account enforcement: limits apply to your account as a whole, not individual API keys. All keys share the same rate limit pool.
• Continuous replenishment: capacity refills steadily over the window period rather than resetting all at once (token bucket model).
• Per-API granularity: each API has its own independent rate limits. WebSocket, Document AI, and LLM APIs have different limits from standard REST APIs. Check your specific API below.

Your current limits are on the Dashboard → Rate Limits page.

Rate limits vary significantly by API type and plan. Review the limits for each API below before building your integration.

Speech to Text limits

Real-time REST

WebSocket Streaming

Batch

Text to Speech limits

Limits are the same for Bulbul v3 and Bulbul v4 Flash.

Real-time REST and HTTP Streaming

WebSocket Streaming

Instant Voice Cloning limits

Translation & Text Services limits

Text translation (POST /translate)

Transliteration and Language Identification share the same limits.

Document Translation

Job lifecycle endpoints (POST /translate/document/jobs, start, live-status, export, export/status):

Language model limits

Default models

Applies to: gemma4, glm5.3, deepseekv4.1-flash-1017

Sarvam-105B model

This large model has lower limits due to its compute requirements.

Applies to: sarvam-105b, sarvam-105b-conversations

Dubbing limits

Document AI limits

Document AI limits are set per operation and scale with your plan. Each limit assumes the number of pages per request shown below.

Exceeding Your Limits

REST rate limits (429 / 503)

When your account exceeds a rate limit, the API returns a 429 Too Many Requests response:

{
  "error": {
    "message": "Rate limit exceeded",
    "code": "rate_limit_exceeded_error"
  }
}

A 503 Service Unavailable can also occur under high backend load. Retry it the same way you would a 429, using a fixed exponential backoff. See Retry with exponential backoff for a reference implementation.

Exhausted credits (402)

A 402 Payment Required is not a rate limit. It means the organisation’s credit balance is empty, and no retry will clear it. Add credits on Settings → Billing and the account resumes immediately, with no key rotation or redeploy. See Billing & payments and Out of credits.

WebSocket connection limits

The limiter reacts to how fast new connections are opened, not how many are held open at once:

• Connections opened in a fast burst can be rejected well below the concurrent ceilings listed above.
• The same number of connections opened gradually (for example, 300ms apart) and held open simultaneously can succeed past those ceilings.

A rejected connection is closed, not queued. The close code is 1003, with this reason text:

Rate limit exceeded. Visit the API Dashboard to review and manage your subscription.

This differs from the 4xxx range described elsewhere on this site for application-specific closures.

Upgrading Your Limits