Price of a request
Each model lists one price per meter inpricing of /v1/models and on its page at naga.ac/models. Most chat models charge per token:
For example,
anthropic/claude-sonnet-5.5 costs "0.000001" credits per input token and "0.000005" per output token. A request with 2,000 input tokens and 500 output tokens costs 2,000 × 0.000001 + 500 × 0.000005 = 0.0045 credits.
Some models have a price range. A row of their pricing then carries max_price next to price, because a provider that serves them costs NagaAI more. NagaAI tries the providers at price first, and you pay the rate of the provider that answered, never below price and never above max_price. To pay only price, send service_tier: "flex" or cap the rate with routing.max_price, see Limit the price.
A chat model that other providers sell too follows the market down. Every ten minutes NagaAI checks their prices for the model. When the cheapest offer falls, price falls to half of it within minutes. A fall of more than 80 % waits for a second check ten minutes later, so one provider’s typo doesn’t reach your bill. max_price is half of the model maker’s own price, and the model has one only while a provider that costs NagaAI more serves it. A price never goes up this way: a rise is announced to paying users in advance. A request pays the price in force when it is served, and GET /v1/models always lists the current one.
Each web search a model runs costs 0.005 credits on top of the tokens, $5 per 1,000 searches, on every paid model. A search that fails costs nothing. See Web search.
A cache write you order costs the cache_write_input_token price, and a one-hour write costs cache_write_1h_input_token, as at OpenRouter. A model without those rows charges a write at its input price. You order a write with cache_control, see Prompt caching. A write the provider or NagaAI’s route made on its own, without cache_control in the request, is reported in usage and costs nothing. Audio and speech models also charge per audio token, per hour of audio or per character. An image model sold per picture, such as Seedream, Recraft or Ideogram, charges per delivered picture (output_image_image, unit: "image"). The price is the same whatever the picture’s size and whichever provider served it. Image models that their maker sells by the token, such as GPT Image and Gemini, charge per image token. Some rerank models charge per search (input_text_search), see Rerank. A video model charges per second, with one price for each resolution and sound setting. Some video models also cost at least a fixed amount per clip. Some models charge a fee per request on top of their units, a request row in pricing. A speech model sold per generation, for example, costs the same whatever the length of its text. On a video model the fee is per clip, and its row carries the resolution and sound it applies to. The fee has a line of its own in the request’s charges. A model can also price one job differently from its others. Such a row carries operation in conditions, for example {"operation": "image_generation"}, and inside that job it replaces the row of the same meter without it. The full list for a model is in its pricing.
Usage in the response
The response carries the token counts that set the price in itsusage object, but no cost. Each format names the counts its own way:
NagaAI charges the cached part of the input at the
cached_input_token price and the rest at the input_token price. NagaAI charges a cache write at its own price, and only when you ordered it. In Responses and Chat Completions the rest is the input count minus the cached and the written counts. In Messages input_tokens already leaves both out, and cache_creation.ephemeral_5m_input_tokens and cache_creation.ephemeral_1h_input_tokens split the write by lifetime. NagaAI returns the audio and image counts only when the provider reports them, and total_tokens can differ from input plus output.
A request with only text in it can’t hold more tokens than it has bytes, plus a little for message roles and tool definitions. When a provider reports more input than that, the extra is instructions of its own, and NagaAI doesn’t charge you for them. The charge lines and Logs show the input NagaAI charged, and usage in the response keeps the count the provider reported. For a request with images, audio, files or earlier reasoning, NagaAI charges the provider’s count.
In a stream, usage arrives at the end. Responses puts it in response.completed and Messages in message_delta. Chat Completions sends it in the last chunk only when you set stream_options: {"include_usage": true}.
To learn what one request cost, call GET /v1/requests/{id} with the same API key after the response ends. Pass the request id from the x-request-id response header, or the id of a chat response such as resp_… or chatcmpl-…. The answer carries the model, the token counts, the cost in credits and one line per unit billed, at the meter’s price. The response itself carries no cost, so this is how your code learns it, for a stream too. A request shows up there a few seconds after it ends and stays readable for 30 days. Only the key that made it can read it, and any other id gets 404.
To see spend per day, model and key, open the dashboard or call /v1/account/activity.
Which requests cost credits
NagaAI charges a request after its answer reaches you, at the price of the model and provider that answered. When a provider fails and NagaAI retries on another one, you pay only for the answer you get. A request that fails on every provider costs nothing. So does a stream that a provider breaks midway. A disconnect doesn’t stop the charge. NagaAI finishes the work after your client goes away and charges the full request. This holds for streams and for image, audio, embedding and moderation requests. Set an output limit on requests you might abandon. An answer cut short bymax_tokens costs only the tokens it used.
Models whose id ends in :free cost nothing.
Payment methods
Top up on the Credits page with Add credits, from $5 to $5,000 at a time. NagaAI adds the credits when the payment completes.
A company can pay a top-up of $1,000 or more by bank transfer against an invoice. Ask for it through the Contact sales form. NagaAI adds the credits on the day the transfer arrives.
On the same page you can set a low balance threshold, and NagaAI emails you when the balance falls below it. Support reviews requests to refund unused credits case by case, within 14 days of the purchase. A refund or chargeback on a card, PayPal or CoinGate payment removes the credits it paid for, which can take the balance below zero.