Understand tokens and budgets¶

The composer shows the current context size; expand its metrics to see Input, Output, Sum and Cost. Completed outputs have a chevron beside the Word and PDF icons that reveals Elapsed, Context, Input and Output below. It starts closed and remembers your choice for later outputs and restarts. Large token values use k; hover for the fuller value and available counting detail. Trace IDs and request counts remain in diagnostics.

Metric	Meaning
Context	The active model request's context, not the sum of every request
Input	Accumulated input usage; repeated context can be counted again on later calls
Output	Accumulated model output usage
Sum	Input plus output
Cost / Est. cost	The current chat's reported charges or labelled token-price estimates; hover for coverage and available account information
Elapsed	Time spent on the displayed turn, including tool work and waits

Provider-native counts are used when supported. Unsupported or unavailable counters remain approximate. Cached-input billing and total input tokens are different concepts; these pills are not a currency invoice. After compaction the active context can shrink while cumulative Input continues to rise.

Provider spending¶

The pill after Sum uses this chat's saved provider connections and runs separately from model generation. Budget observations refresh at most once per minute per connection; their observation time appears on hover.

Cost shows the current chat's reported model-request charges. It includes Main, Coder and auxiliary calls in the turn accounting. OpenRouter supplies usage.cost; xAI supplies usage.cost_in_usd_ticks. Other compatible servers can supply an amount with its currency. Repeated streaming usage frames are counted once per response.
Est. cost and ~ identify a total that includes estimates. For supported direct OpenAI models, Nawa uses the request's model, processing tier, cached input, cache writes and long-context pricing. A reported monetary charge takes precedence, including a reported zero. Hover shows the pricing source, check date and assumptions. These estimates exclude tool fees, taxes and account discounts.
(partial) means some requests have neither a reported charge nor supported pricing. They are excluded from the displayed amount; hover states how many.
The hover also shows API key usage and account balances when the provider exposes them, including the Coder's configured connection. OpenRouter reports all-time key usage, a spending cap and remaining key allowance. These cover other chats or apps using the same key. DeepSeek reports account balances in their original currency. Neither is substituted for the current chat's cost.
— means monetary information is unavailable. It does not mean the service is free or the balance is zero. Older chat turns without saved cost observations remain unreported.

OpenAI and Anthropic organization cost reporting requires separate administrative access. Gemini uses its billing/usage services. Nawa does not request organization-wide credentials. Token-price estimates are separate from account balances. The provider catalogue records explicitly supported model prices and their source; unknown models, local servers and custom endpoints remain unavailable unless they report charges. Older requests without saved pricing records are not repriced using the chat's current model.

When a provider rejects a request, Nawa displays its reason and any stated retry interval. A spending rejection does not undo earlier completed actions. Send Continue after resolving the limit or waiting the provider's stated interval. Nawa cannot continue a request the provider refuses.

OpenRouter's in-flight spending limit is a temporary allowance for running and recently completed requests. It can reject a large Coder request while the balance is positive and a smaller follow-up answer still succeeds. This is separate from context size, the API key's spending cap and account balance.

Controls¶

In Settings → Budgets, Output tokens per model call limits an individual reply; leave it blank for the native allowance. A smaller output budget may cut a useful reply short.

Max parallel agents limits simultaneous agents. Progress timeout is renewed by authoritative operation or workspace progress; it is not a fixed total duration for productive work. Document controls are explicitly labelled in characters and govern extraction, not model token accounting.

Each model's context size, and the percentage at which Nawa compacts it, are shown beside the model in Settings → LLM as Compact at (%). Compaction makes room by condensing older context and retaining access to the transcript and stored evidence. It is not a reset of the task or its usage totals. See Context and compaction.