> ## Documentation Index
> Fetch the complete documentation index at: https://help.casaconect.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# AI credits
> How AI usage is metered and paid for: what consumes credits, what each model costs, and how to control spending with auto-refill, budgets and per-member limits.
Everything the agent does with AI is paid for with **AI credits**: prepaid funds that belong to your organization. The principle is simple, and it is stated in the product:
> Credits pay for the agent's work: reading documents, drafting, and browsing the web. You pay what the AI providers charge us, with EU hosting adding 10% on Claude models.
Model usage is billed at the provider's list price, with **no margin** added by Casa Conect. The 10% on Claude models is what Amazon Bedrock charges for keeping inference inside EU regions. It is the price of EU processing, passed through unchanged.
## Buy credits
In **Settings → Access & billing**, choose a pack or enter a custom amount.
| Pack | |
| :----- | :----------------------------- |
| €25 | |
| €100 | |
| €500 | |
| Custom | Any amount from €10 to €10,000 |
Credits are charged to the organization's saved card and are available immediately. The page shows your balance and an estimate: *"About N days left at the current rate."*
When credits run out, *AI features pause until you top up*. Conversations, workflows and indexing stop, including work the agent was doing unattended. Your data is unaffected. Use **auto-refill** to avoid this.
## Auto-refill
*Charge the saved card for a credit pack whenever the balance drops below the threshold.* Choose the threshold and the pack.
Two safeguards protect you from a runaway charge:
* At most **five automatic charges per day**. If the limit is reached, auto-refill **pauses** until an admin switches it on again.
* If the card is declined, auto-refill pauses and the page tells you so.
## Control spending
An *optional cap on the organization's AI spend per calendar month, in USD*. When it is reached, AI work stops until the next month, or until an admin raises the cap.
See *what each member has spent this month*, and set a monthly limit for any of them. A member who reaches their limit is paused. The rest of the team is not.
Block the most expensive models for everyone.
Before any model is called, Casa Conect checks, in this order: an active subscription, the [model policy](/models/model-policies), the credit balance, then the organization budget and the member's limit. A refusal at any step costs nothing, and the member is told which one applied.
## What consumes credits
Usage is broken down by activity under **Settings → Access & billing → Usage**:
| Activity | What it covers |
| :-------------------------------------------- | :--------------------------------------------------- |
| **Agent conversations** | The model you chat with. Normally the largest share. |
| **Agent helpers working in the background** | Read-only helpers the agent starts |
| **Naming conversations** | The short title written for each conversation |
| **Filling in templates** | The [Populate](/workflows/populate) workflow |
| **Anonymizing templates** | The [Anonymize](/workflows/anonymize) workflow |
| **Shared space summaries** | Extracting summary values from a space's documents |
| **Applying matter templates** | Restructuring a matter's description |
| **Searching your documents** | Query understanding and reranking |
| **Reading and indexing new documents** | Parsing and embedding uploads |
| **Remembering context between conversations** | Conversation memory |
| **Voice dictation** | Speech to text, billed per second |
| **Browsing the web** | Cloud browser time, billed per second |
The same page breaks usage down **By model**, showing calls, tokens and cost, and **by member**.
## What each model costs
Model usage is measured in **tokens**, which are pieces of words. As a rule of thumb, 1,000 tokens is about 750 English words, and somewhat fewer in Spanish. **Input** is everything the model reads: your message, the conversation so far, and the documents and search results it looked at. **Output** is what it writes. Output costs more than input.
Provider list prices in US dollars per million tokens, as of **September 2026**:
| Model | Input | Cached input | Output |
| :-------------------------- | ------: | -----------: | ------: |
| ChatGPT 5.6 Luna (**Auto**) | \$0.20 | \$0.02 | \$1.20 |
| ChatGPT 5.6 Terra | \$2.00 | \$0.20 | \$12.00 |
| Sonnet 5 EU | \$2.20 | \$0.22 | \$11.00 |
| ChatGPT 5.6 Sol | \$4.00 | \$0.40 | \$20.00 |
| Opus 5 EU | \$5.50 | \$0.55 | \$27.50 |
| ChatGPT 6 Astra | \$10.00 | \$1.00 | \$50.00 |
EU Includes Amazon Bedrock's 10% premium for EU regional inference.
Providers change their prices, and Casa Conect follows them. Your **Usage** page always shows what you were actually charged. This table is a guide.
### A sense of scale
| Task | Roughly |
| :---------------------------------------------- | :--------------------------------------------------------------------------------------------- |
| A quick question on Auto | A fraction of a cent |
| Summarizing a 30-page contract on Sonnet 5 | A few cents |
| A long drafting and redlining session on Opus 5 | Tens of cents to a few dollars |
| Reading and indexing a 100-page document | About 40 cents. A little over a dollar if poor scan quality calls for the more capable parser. |
**Cached input** is why long conversations cost less than you might expect. When the model re-reads material it saw moments ago, the provider charges about a tenth of the normal price.
### Long conversations
On the ChatGPT frontier models, a single request that contains more than **272,000 input tokens**, which is several hundred pages, is priced by the provider at twice the input rate and one and a half times the output rate. In practice this affects only very long conversations over very large documents. Starting a new conversation for a new subject avoids it.
## Other usage
| Service | Basis |
| :--------------- | :---------------------------------------------------------------- |
| Document parsing | Per page. Poor scans that need the more capable parser cost more. |
| Search reranking | Per search |
| Voice dictation | Per second of audio |
| Browser sessions | Per second of browser time |
## Keeping costs down
* **Leave Auto on** for everyday work, and pick a stronger model only when the task needs it. See [Choosing a model](/models/choosing-a-model).
* **One subject per conversation.** The whole conversation is re-read on every turn.
* **Reference documents** instead of pasting long text. The agent then reads only the sections it needs.
* **Use workflows** for template filling. They use an economical model wherever that is sufficient.
* **Close browser sessions** when you have finished with them.
* **Upload clean scans.** Poor scans cost more to read and give worse results.