# Paying for your own inference or answering on visitors' AI

Source: https://tensorpass.app/compare/paying-for-your-own-inference
Published on 2026-11-12.

> A site paying its AI provider on its own key compared with AI Connect, where the visitor's own AI account answers, on cost, keys, models, tools and phones.

A site that adds an AI chat today pays its provider for every token, on an API key its server holds. That bill is the site's **[inference cost](/glossary/inference-cost)**: the price per million input and output tokens times the tokens each answer uses. AI Connect moves the bill to the visitor. The visitor's own AI account answers through **[AI Wallet](/glossary/ai-wallet)**, a browser extension that keeps their accounts in an encrypted vault on their computer, and the site pays no inference.

## Eight points side by side [#eight-points-side-by-side]

| Point | Site's own key | Visitors' AI through AI Connect |
| --- | --- | --- |
| Who pays | The site, per token, at its provider's price | The visitor's plan, credits or API key |
| Who holds the keys | The site's server, with its own key | The wallet, with the visitor's credentials; the site holds none |
| Model choice | The site's pick, one model per feature | The visitor's pick of account and model for each site, out of 20 kinds of account |
| Tools and MCP | Any tool the server runs, including a provider's hosted tools | Function tools from the site's MCP server, called from the page |
| What the visitor needs | A browser | AI Wallet in Chrome 140 or later, and an AI account |
| Phones | Every device | Chrome on a computer only; a mobile app designed and unbuilt |
| Floods and limits | Every request on the site's bill, with rate limits the site builds | Each answer on the asking visitor's account; a default cap of 2,000,000 tokens a day per site in the wallet |
| Status on 16 November 2026 | Available from every provider | 0.1.0 on npm; wallet in the Chrome Web Store's review |

## What one answer costs on a site's key [#what-one-answer-costs-on-a-sites-key]

One answer costs a site between $0.00056 and $0.038 at the list prices of November 2026, by its size and its model. A university admissions chatbot averaged 1,748 input tokens and 60 output tokens per message. A **retrieval augmented generation** (RAG) query, where the site adds passages from its own content to the prompt, averaged 3,865 input and 426 output tokens in its plain form. Its iterative form, which retrieves several times before answering, averaged 12,814 input and 1,207 output tokens.

| Workload | Tokens in and out | Model, price per million in and out | One answer | 100,000 answers |
| --- | --- | --- | --- | --- |
| Admissions chat message | 1,748 and 60 | gpt-5-mini, $0.25 and $2 | $0.000557 | $55.70 |
| Plain RAG query | 3,865 and 426 | Claude Haiku 5.5, $0.10 and $0.50 | $0.000600 | $59.95 |
| Plain RAG query | 3,865 and 426 | Claude Sonnet 4.6, $3 and $15 | $0.0180 | $1,798.50 |
| Iterative RAG query | 12,814 and 1,207 | gpt-6.1-sol, $2 and $10 | $0.0377 | $3,769.80 |

Each line multiplies the tokens by the price per million and divides by one million. On the last line, 12,814 input tokens at $2 per million cost $0.025628 and 1,207 output tokens at $10 per million cost $0.01207, for $0.037698 an answer. The admissions chatbot itself cost $11.58 over two weeks on GPT-4o mini. The [study of one answer's cost](/blog/cost-of-one-ai-answer) shows the arithmetic for more models.

Enterprise spending on AI deployment reached $176.8 billion in 2026, 5.4 times the year before, with $25.0 billion on chatbots. Average spend per company came out at 4.7 times the prior year, where companies had planned 1.9 times. These figures come from about 400 US decision makers surveyed in September 2026.

## What changes for the site [#what-changes-for-the-site]

**Cost.** The inference line of an AI feature goes to $0 on AI Connect, since each answer runs on the account of the visitor who asked. The site still pays for its MCP server and the search or database behind its tools.

**Code.** A site that calls a model through the OpenAI SDK, the Anthropic SDK, Google Gen AI, the Vercel AI SDK, LangChain or LlamaIndex keeps that code and passes `aiconnect.fetch` to the SDK. The call moves from the server route into the page, and the route's data work stays on the server as tools on an **[MCP](/glossary/model-context-protocol)** server (Model Context Protocol, the open protocol for serving tools to a model). The [guide for existing SDKs](/docs/guides/existing-sdks) shows the change for each SDK.

**Model.** The site gives up the choice of model. The visitor gives each site an account and a model in the wallet, and the site receives the model's name. Some accounts drop request options: the ChatGPT plan leaves out `maxTokens`, `temperature`, `topP` and `stop`, and the Claude plan and Claude API key leave out `temperature` and `topP`. The wallet passes on function tools only, so a provider's hosted web search or code execution stays out.

**Abuse.** On its own key, a site pays for every request a script sends to its chat. On AI Connect, the wallet caps each site at 2,000,000 tokens a day by default, limits the chats it starts per minute and runs at once, and answers a site past a limit with a 429 before any provider receives the call. Guards in the page check each question and each answer for **[prompt injection](/glossary/prompt-injection)** (text planted in data to steer the model), as the [guards page](/docs/modules/guards) describes.

## When does the site's own key fit better? [#when-does-the-sites-own-key-fit-better]

A site's own key fits these cases better than visitors' AI on 16 November 2026:

- **Visitors without an AI account.** A visitor with no plan, credits or API key answers through AI Connect once they open an account, such as OpenRouter's or the TensorPass plan, and the site's key answers everyone at once.
- **Phone traffic.** AI Wallet runs in Chrome on a computer, so a phone visitor answers through a sign in, and a site that answers every phone visitor without one needs its own key.
- **Content made once for every reader.** A product description or a page summary generated once and cached serves every visitor for the price of one call, where visitors' accounts would each repeat the work.
- **One fixed model.** A feature tuned for one model, a fixed temperature or a provider's hosted tools needs the site's key, since the visitor picks the model on AI Connect.
- **A launch before the stores.** AI Wallet waits on the Chrome Web Store's review, and until it ends a visitor installs it by hand from https://tensorpass.app/wallet.

## Run both on one site [#run-both-on-one-site]

A site keeps its server route for visitors without the wallet and sends connected visitors through `aiconnect.fetch`. A call through `aiconnect.fetch` from a visitor who has not connected fails with status 401. On that status the site answers through its own route. Connected visitors get their answers from their own account. The site's bill covers the others. The [comparison with other bring your own AI tools](/compare/other-bring-your-own-ai-tools) covers the ways a visitor's account reaches a site, and the [bring your own AI entry](/glossary/bring-your-own-ai) defines the term.

## Sources [#sources]

- [OpenAI, API pricing, read 12 November 2026](https://developers.openai.com/api/docs/pricing)
- [Anthropic, Pricing, read 12 November 2026](https://platform.claude.com/docs/en/about-claude/pricing)
- [arXiv 2507.11272, cost and token use of a university admissions chatbot, 15 July 2025](https://arxiv.org/abs/2507.11272)
- [arXiv 2409.02361, token use of plain and iterative RAG, 2024](https://arxiv.org/pdf/2409.02361)
- [Menlo Ventures, 2026: the state of enterprise AI, 3 November 2026](https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/)
- [TensorPass, Use the AI SDK you already have, 12 November 2026](/docs/guides/existing-sdks)
- [TensorPass, Providers, 12 November 2026](/docs/providers)
- [TensorPass, Guards, 12 November 2026](/docs/modules/guards)
- [TensorPass, Changelog, 0.1.0, 16 November 2026](/docs/changelog)

## Questions

**How much does one AI answer cost a site on its own key?**
A chat message of 1,748 input and 60 output tokens costs $0.000557 on gpt-5-mini at $0.25 and $2 per million tokens. A retrieval query of 12,814 input and 1,207 output tokens costs $0.0377 on gpt-6.1-sol at $2 and $10 per million, or $3,769.80 for 100,000 queries.

**Does AI Connect remove every cost of a site's AI feature?**
It removes the inference bill, since the visitor's account pays each answer. The site still runs its MCP server and the search or database behind its tools, and pays for that hosting.

**Which visitors get no answer through AI Connect?**
On 16 November 2026, a visitor who holds no AI plan, credits or API key and opens none gets no answer through it. A visitor without AI Wallet, on a computer or a phone, answers through a sign in with OpenRouter, or with the TensorPass plan on a site that passes a site key, and can open either account in that window. AI Wallet runs in Chrome 140 or later on a computer.

**Does the site still choose the model?**
The visitor chooses the account and model for each site in the wallet, and the site receives the model's name in its connection state. The ChatGPT plan leaves out temperature, topP, maxTokens and stop, and the Claude accounts leave out temperature and topP.

**Does a site have to give up its own key?**
No. A site that calls a model through an AI SDK passes aiconnect.fetch for connected visitors. A call from a visitor who has not connected fails with status 401, and on that status the site answers the visitor through its own server route.