How chat tokens and credits work

Each Argon reply uses tokens, and TryCarbon AI settles that usage in credits. Expand Recent usage in chat to see input tokens, output tokens including thinking, and the credits charged. Your account credit history records the related balance changes.

What counts as input and output?

Input includes your new question and the earlier conversation sent with it. It is not just the number of words in your latest message. Output includes the answer and any billable thinking tokens reported by the provider.

Cached input is part of total input, not an extra quantity to add again. Where Google reports cached tokens, we account for the different rate. Google's token documentation explains the provider's usage fields; the receipt in this website reports what your request actually used.

A real small request

During our October 29, 2026 connection check, the prompt “Reply with exactly: Argon connection verified.” returned “Argon connection verified.” The API reported 39 input tokens and 9 output tokens, for 48 total. Those counts show why token usage should come from the API rather than a word-count estimate. This was a connectivity test, not a quality benchmark or a typical usage forecast.

Your counts will vary with the prompt, chat history and response. A much longer answer may also use thinking tokens that are not visible as prose.

See the rate before you buy

The Pricing rate table shows credits per 1,000 uncached input, cached input and output tokens, including thinking. At the rate available through November 30, 2026, 1,000 uncached input and 500 output tokens cost 25 credits after whole-turn rounding. The verified 39-input/9-output request above cost 5 credits. These are examples, not a promise about message length. The scheduled December 1 rates are shown on the same page.

No prepayment or frozen allowance

Your balance stays unchanged while a reply is generated. The server counts the input, calculates an output limit that fits the current balance, and checks provider-reported usage during streaming. If the allowance is exhausted, generation stops; only confirmed actual usage is charged after the reply ends. Thinking tokens share the output limit.

There is no fixed 150-credit entry requirement. A balance of 100 credits can fund a shorter reply. Very long input can itself exceed a small balance; start a new conversation or shorten the context in that case. An extremely small output budget can be consumed by thinking before a visible answer is produced.

Whole-turn charges round upward in steps of 5 credits up to 100, 50 credits up to 1,000, and 100 credits above that. The minimum non-zero charge is 5 credits. A provider overrun is absorbed by the site rather than driving your balance negative. Only one generation can run per account at a time.

A confirmed rejection is not charged. An interrupted reply without final provider usage remains pending, with no frozen credits; we do not invent a charge from partial text. Confirmed search or image costs are still settled even when another part of the turn is pending. Several such interruptions in a short period can temporarily pause new requests while usage is checked. A completed usage receipt states the actual settled charge.

Search and image costs use the same balance

Google web search costs 46.2 credits per executed, nonempty unique query, before whole-turn rounding. The same query is counted once within a provider request; a new request can incur a new charge. A message can trigger multiple queries. We apply this paid search rate regardless of any upstream free quota. Search fees and text-token costs are combined before rounding: the 1,000-input/500-output example plus one query costs 70 credits at the introductory Argon rate.

The image tool uses Gemini 3.1 Flash Lite Image for 1K square images. Image input, text/thinking output and image output are metered separately, with each image-generation cost rounded upward to the next 10 credits. One measured image used 46 input tokens, 378 text/thinking output tokens and 1,120 image-output tokens, costing 120 credits for that image call. Argon chat tokens used to call the tool are additional; this is not a fixed price per image.

Stopping a reply does not reverse search or image work already executed. A generation budget check can reject an image call before it starts if the remaining balance cannot cover its output ceiling; it does not freeze credits. See all model rates and package features.

Choose a context window

Choose 128K, 256K, 512K or 1M beneath the composer (default 128K). This is a capacity limit, not a prepaid amount. The server counts the history, instructions and new message before generation. If they exceed the chosen window, the conversation asks you to start a New Chat. Your old history remains saved. We do not automatically compress, summarize or silently discard messages.

What a package gives you

The packages contain 10,000 or 20,000 credits. They are one-time purchases, do not renew and do not expire. There is no fixed “number of messages” promise because message length and context vary. Registration does not include free inference credits.

A future model may have a different rate. Carbon is not currently available here, and buying a package does not promise Carbon access. Check Pricing for the purchase amount and Carbon status for model availability.

Keep longer projects manageable

Ask for the detail you need, rather than a long answer by default. Use a new conversation for a new subject. When carrying a project into a fresh chat, paste a concise summary and the exact source passages that matter. These choices reduce unnecessary context while leaving you in control of the task.

Open chat and inspect your usage, or choose a credit package.