# The cost of one AI answer on a website's own API key

Source: https://tensorpass.app/blog/cost-of-one-ai-answer
Published on 2026-11-12, updated on 2026-11-16.

> One chat answer costs a site $0.0002 to $0.0082 on its own key and one RAG answer $0.0006 to $0.024, at list prices of 12 November 2026. The arithmetic.

On a site's own API key, one chat answer costs between $0.0002048 and $0.008192 at the list prices of 12 November 2026, and one answer from a retrieval chatbot costs between $0.0005995 and $0.02398. The spread comes from the model: the cheapest and the dearest of the eight models below differ by a factor of 40. At 1,000,000 answers, the same chat costs the site $204.80 on Claude Haiku 5.5 and $8,192 on Claude Opus 5.5. This article shows the arithmetic, states its assumptions, and gives what the site pays when the visitor's own AI answers through AI Connect.

## The formula per answer [#the-formula-per-answer]

The **inference cost** of one answer is the price of the tokens the model reads plus the price of the tokens it writes ([inference cost](/glossary/inference-cost) defines the term). Providers list both prices in US dollars per million tokens.

cost per answer = tokens in × price in ÷ 1,000,000 + tokens out × price out ÷ 1,000,000

A chat answer on Claude Sonnet 5.5 works out as 1,748 × $2 ÷ 1,000,000 = $0.003496 for the input, plus 60 × $10 ÷ 1,000,000 = $0.000600 for the output. The answer costs $0.004096. Ten thousand answers cost $40.96, and a million cost $4,096.

## Assumptions behind the tables [#assumptions-behind-the-tables]

- **Chat answer.** 1,748 input tokens and 60 output tokens, the average per message of a university admissions chatbot measured in production. That deployment cost $11.58 over two weeks on GPT-4o mini.
- **RAG answer.** **Retrieval augmented generation** (RAG) adds passages found in the site's content to the question before the model answers. A plain RAG pipeline averaged 3,865 input and 426 output tokens per query on a question answering benchmark, and an iterative pipeline, which retrieves several times, averaged 12,814 and 1,207.
- **Prices.** The list prices of 12 November 2026 at the standard rate, with no prompt caching discount, no batch discount and no free tier.
- **Token counts.** The same counts apply to every model. Each model's tokenizer splits the same text into its own number of tokens, and a site with longer instructions or more tool definitions sends more input per answer.
- **Price tiers.** Claude Haiku 5.5 costs $0.10 and $0.50 for prompts up to 100,000 tokens; every prompt here stays under 13,000. `gpt-6-luna` carries the same prices as Claude Haiku 5.5, so its costs match that row.

## One chat answer [#one-chat-answer]

| Model | Tokens in | Tokens out | Price in per 1M | Price out per 1M | Cost per answer | Per 10,000 answers | Per 1,000,000 answers |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Claude Haiku 5.5 | 1,748 | 60 | $0.10 | $0.50 | $0.0002048 | $2.05 | $204.80 |
| `gpt-5-mini` | 1,748 | 60 | $0.25 | $2.00 | $0.000557 | $5.57 | $557.00 |
| `gemini-3.5-flash-lite` | 1,748 | 60 | $0.30 | $2.50 | $0.0006744 | $6.74 | $674.40 |
| Mistral Large 4 | 1,748 | 60 | $0.70 | $2.10 | $0.0013496 | $13.50 | $1,349.60 |
| `gemini-3.8-flash` | 1,748 | 60 | $0.75 | $3.75 | $0.001536 | $15.36 | $1,536.00 |
| `gpt-6.1-sol` | 1,748 | 60 | $2.00 | $10.00 | $0.004096 | $40.96 | $4,096.00 |
| Claude Sonnet 5.5 | 1,748 | 60 | $2.00 | $10.00 | $0.004096 | $40.96 | $4,096.00 |
| Claude Opus 5.5 | 1,748 | 60 | $4.00 | $20.00 | $0.008192 | $81.92 | $8,192.00 |

Input carries most of the bill for a chat answer. On Claude Sonnet 5.5 the 1,748 input tokens cost $0.003496 and the 60 output tokens $0.0006, so input makes 85% of the cost.

## One RAG answer [#one-rag-answer]

| Model | Tokens in | Tokens out | Price in per 1M | Price out per 1M | Cost per answer | Per 10,000 answers | Per 1,000,000 answers |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Claude Haiku 5.5 | 3,865 | 426 | $0.10 | $0.50 | $0.0005995 | $6.00 | $599.50 |
| `gpt-5-mini` | 3,865 | 426 | $0.25 | $2.00 | $0.00181825 | $18.18 | $1,818.25 |
| `gemini-3.5-flash-lite` | 3,865 | 426 | $0.30 | $2.50 | $0.0022245 | $22.25 | $2,224.50 |
| Mistral Large 4 | 3,865 | 426 | $0.70 | $2.10 | $0.0036001 | $36.00 | $3,600.10 |
| `gemini-3.8-flash` | 3,865 | 426 | $0.75 | $3.75 | $0.00449625 | $44.96 | $4,496.25 |
| `gpt-6.1-sol` | 3,865 | 426 | $2.00 | $10.00 | $0.01199 | $119.90 | $11,990.00 |
| Claude Sonnet 5.5 | 3,865 | 426 | $2.00 | $10.00 | $0.01199 | $119.90 | $11,990.00 |
| Claude Opus 5.5 | 3,865 | 426 | $4.00 | $20.00 | $0.02398 | $239.80 | $23,980.00 |

A RAG answer costs about 3 times a chat answer on the same model: $0.01199 against $0.004096 on Claude Sonnet 5.5. The RAG input is 2.2 times the chat input, and the RAG answer is 7.1 times longer.

### The iterative pipeline [#the-iterative-pipeline]

| Model | Tokens in | Tokens out | Price in per 1M | Price out per 1M | Cost per answer | Per 10,000 answers | Per 1,000,000 answers |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Claude Haiku 5.5 | 12,814 | 1,207 | $0.10 | $0.50 | $0.0018849 | $18.85 | $1,884.90 |
| `gemini-3.8-flash` | 12,814 | 1,207 | $0.75 | $3.75 | $0.01413675 | $141.37 | $14,136.75 |
| Claude Sonnet 5.5 | 12,814 | 1,207 | $2.00 | $10.00 | $0.037698 | $376.98 | $37,698.00 |
| Claude Opus 5.5 | 12,814 | 1,207 | $4.00 | $20.00 | $0.075396 | $753.96 | $75,396.00 |

The iterative pipeline costs 3.1 times the plain one on Claude Sonnet 5.5, $0.037698 against $0.01199 per answer.

### Prices dated to change [#prices-dated-to-change]

`gemini-3.8-flash` moves from $0.75 and $3.75 to $1.50 and $7.50 on 1 January 2027. From that date a chat answer costs $0.003072 on it and a RAG answer $0.0089925, twice the figures above.

## Who pays with AI Connect [#who-pays-with-ai-connect]

With AI Connect, the site's inference line is $0 for every row of the three tables. The site's page sends the conversation to AI Wallet, the extension in the visitor's browser, and the wallet calls the provider with the visitor's own credential: a plan sign in such as ChatGPT or Gemini, or an API key. The provider counts the answer against that account. A visitor on a plan uses the room the plan leaves, and a visitor on a key pays the list prices above. This is the model of [bring your own AI](/glossary/bring-your-own-ai).

The wallet shows the room left on each plan or key, and its Auto choice moves to the next account when one runs out of room. The site keeps its measure of each answer: it receives the token counts the provider reported, the model that answered and the finish reason. A site already built on an AI SDK moves its calls to the wallet with `aiconnect.fetch`, as the [AI SDK guide](/docs/guides/existing-sdks) shows. The packages of version 0.1.0 are on npm since 16 November 2026 ([AI Connect 0.1.0 is on npm](/blog/aiconnect-on-npm)).

## Costs the site keeps [#costs-the-site-keeps]

The site keeps every cost outside the model's tokens.

| Cost line | On the site's own key | With AI Connect |
| --- | --- | --- |
| Model tokens per answer | The tables above | $0 for the site; the visitor's plan or key answers |
| Pages and their hosting | Paid by the site | Paid by the site |
| MCP server | Paid by the site | Paid by the site |
| Search index of the site's content | Paid by the site | Paid by the site |

The **MCP server** exposes the site's tools, such as a catalogue search or a booking check, over the [Model Context Protocol](/glossary/model-context-protocol), and the visitor's model calls them from the page. Each call runs on the site's server, so the server needs bounds and a rate limit, as the [MCP guide](/docs/guides/mcp-tools) describes. The free `aiconnect-rag-chatbot` skill builds a keyword index of the site's pages and serves it as two tools, so a RAG chatbot on AI Connect leaves the site with the cost of the index and the search. The comparison [paying for your own inference](/compare/paying-for-your-own-inference) sets the two models side by side.

## Sources [#sources]

- [OpenAI, API pricing, read 12 November 2026](https://developers.openai.com/api/docs/pricing)
- [Anthropic, Pricing, Claude platform documentation, read 12 November 2026](https://platform.claude.com/docs/en/about-claude/pricing)
- [Google, Gemini API pricing, Google AI Studio, read 12 November 2026](https://aistudio.google.com/docs/pricing)
- [Mistral AI, Mistral Large 4, model documentation, 5 November 2026](https://docs.mistral.ai/models/mistral-large-4-0)
- [arXiv, An Empirical Study of Multi-Agent RAG for Real-World University Admissions Counseling, 15 July 2025](https://arxiv.org/abs/2507.11272)
- [arXiv, Diversify-verify-adapt: Efficient and Robust Retrieval-Augmented Ambiguous Question Answering, 4 September 2024](https://arxiv.org/abs/2409.02361)
- [TensorPass, Introduction, AI Connect documentation, 12 November 2026](https://tensorpass.app/docs)
- [TensorPass, Agent Skills index with the aiconnect-rag-chatbot skill, 12 November 2026](https://tensorpass.app/.well-known/agent-skills/index.json)