Skip to main content
Early accessGnani Evon v3.5 is now available as an API to early access registrants.

What a token is

The model reads and writes text in tokens, pieces of words. In English a token is about four characters, or three-quarters of a word. Text in Indian scripts can take more tokens than English for the same meaning. Every response reports the exact count in usage.

Context window by plan

The context window is the input and the output of one request together. A request over the context window is refused with 400 context_length_exceeded.

Free and Pay-as-you-go plans

There is no tokens-per-minute limit. Enterprise plans have a 128K-token context window. If your account belongs to an organisation, your organisation has its own limits. Contact support to find out what they are or to change them.

How the limits work

Context window and input

The context window is everything the model handles for one request: the whole conversation you send, plus the reply it may write. A request is refused before the model runs when:
  • the input is over the input limit for your plan (62K tokens on Free and Pay-as-you-go), or
  • the input plus max_tokens is over your plan’s context window (64K tokens on Free and Pay-as-you-go, 128K on Enterprise).
The input is estimated from the size of the request body (about 4 bytes per token) before the model runs, so a request close to the limit can be refused even if its exact count would fit. The error message gives the estimate:
In long conversations, drop or summarise the oldest messages before the conversation reaches the limit. Keep the system message.

Reply length

A request without max_tokens can use up to 1,024 output tokens. Asking for more is refused:
Thinking counts toward this limit. If replies come back empty with finish_reason: "length", turn thinking off with "reasoning_effort": "none".

Request rate and concurrency

You can send one request per second, with short bursts of two, up to 60 a minute, and one request can run at a time. Wait for each reply, streamed or not, to finish before you send the next request. Requests that come too fast are refused by the API gateway:
Wait, then retry with backoff. When the response has a Retry-After header, wait at least that many seconds. OpenAI’s SDKs do this for you.

Daily token quota

Every request’s usage.total_tokens (input plus output, thinking included) counts toward 100,000 tokens a day. The count resets at midnight IST. Once it is used up, every request is refused until then, and Retry-After gives the seconds until the reset:
Because each request carries the whole conversation, long chats use the quota faster with every turn.

Estimate the size of a request

Paste text to see how the pre-check above would size it. This is an estimate from the request body size, not the model’s exact token count.

Ways to use fewer tokens

  • Turn thinking off for simple tasks ("reasoning_effort": "none"). Thinking is often the largest part of the output.
  • Trim the conversation. Send only the turns the model needs.
  • Keep the system message short. It is sent, and counted, with every request.
  • Set max_tokens to the length you need.
  • Use web search only when needed. The search results are added to the input.

Other limits


Errors

Error codes, and how to retry with backoff.

Thinking

Thinking counts toward max_tokens and the quota.

Web search

Daily searches and how they use tokens.

Safety and privacy

The 1 MB of text that can be screened.