Running a support workflow (classify, retrieve and answer, email reply) costs far less on GPT-6 Luna than on Claude Fable 5.1, based on list prices as of 2026-10-11 (USD). On the RAG step alone, Luna costs $0.54 per 1,000 tasks and Fable costs $136.54. The gap comes from model choice, token counts and reasoning tokens, not from ticket volume, and the retrieval-based answer is the most expensive step per ticket.
What I counted
A typical bot does more than one thing per ticket. I looked at three steps for each model:
- Classify the ticket.
- Answer from five retrieved documents (RAG).
- Reply to the customer by email.
I report each step separately and do not add them into a single bundle. Real systems usually route easy tickets to a cheaper path and send only hard ones through retrieval, so the right mix depends on your traffic. Chat answers are a separate task, listed below.
One detail matters when comparing steps. My classification task covers 100 tickets per run, so its cost is per 100 tickets. The email and RAG tasks are one ticket per run.
Some figures are measured: one run per task at default settings via OpenRouter, output capped at 16,000 tokens. The rest are estimated. GPT-6 Luna, Claude Haiku 5.5, GPT-6.1 Sol and Claude Sonnet 5.5 are measured. GPT-5.4 Nano, Claude Opus 5.5 and Claude Fable 5.1 are estimated. See the methodology for how I label each.
Reasoning tokens are billed as output tokens. Several cheap models produce thousands of reasoning tokens on a simple classification, so their real cost can sit well above what the list price suggests. GLM 5.3 Flash, for example, used 7,411 reasoning tokens on the classification task.
Cost by step and model
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GPT-6 LunaOpenAI | 0.05¢$0.54 / 1,000 | 1443 | 3,690 / 338 | measured best value | |
| 2 |  GLM 5.3 FlashZ.ai | 0.06¢$0.57 / 1,000 | 1474 | 3,714 / 1,971 | measured best value | |
| 3 |  DeepSeek V4.1 FlashDeepSeek | 0.08¢$0.81 / 1,000 | 1475 | 3,696 / 3,750 | measured best value | |
| 4 | GPT-5.4 NanoOpenAI | 0.12¢$1.16 / 1,000 | 1401 | 3,690 / 338 | estimated | |
| 5 |  Claude Haiku 5.5Anthropic | 0.14¢$1.44 / 1,000 | - | 5,184 / 1,849 | measured | |
| 6 |  Qwen3.8 FlashAlibaba Qwen | 0.19¢$1.92 / 1,000 | - | 3,751 / 2,886 | measured | |
| 7 |  Mistral Small 4Mistral | 0.25¢$2.48 / 1,000 | - | 3,772 / 3,191 | estimated | |
| 8 |  Llama 4 MaverickMeta | 0.26¢$2.58 / 1,000 | 1327 | 3,721 / 2,878 | estimated |
The table below puts the three steps side by side. Email and RAG are per 1,000 tasks. Classification is per 1,000 runs, and each run covers 100 tickets.
| Model | Classify (per 1,000 runs of 100 tickets) | Email reply (per 1,000) | RAG answer (per 1,000) |
|---|---|---|---|
| GPT-6 Luna | $0.64 | $0.07 | $0.54 |
| GPT-5.4 Nano | $1.55 | $0.17 | $1.16 |
| Claude Haiku 5.5 | $0.92 | $0.35 | $1.44 |
| GPT-6.1 Sol | $7.71 | $1.78 | $11.80 |
| Claude Sonnet 5.5 | $8.90 | $3.97 | $27.31 |
| Claude Opus 5.5 | $17.81 | $7.95 | $54.62 |
| Claude Fable 5.1 | $44.52 | $19.87 | $136.54 |
GPT-6 Luna is the cheapest model in this set on all three steps. Claude Fable 5.1 is the most expensive in this table on all three. Other models can cost more still on some steps. Kimi K3, for example, costs more than Fable on classification (11¢ against 4.5¢ per run) and on email replies (5.3¢ against 2.0¢), though not on RAG.
Chat answers follow the same pattern. Per 1,000 chat answers, GPT-6 Luna costs $0.59, Claude Haiku 5.5 costs $0.66, Claude Sonnet 5.5 costs $12.68 and Claude Opus 5.5 costs $25.36.
Why the same list price can cost more
List price explains part of the spread. GPT-6 Luna and Claude Haiku 5.5 both list at $0.10 input and $0.50 output per million tokens, and their step costs are close. Claude Sonnet 5.5 lists at $2.00 and $10.00, and Claude Opus 5.5 at $4.00 and $20.00.
But GPT-6.1 Sol and Claude Sonnet 5.5 have identical list prices ($2.00 input, $10.00 output), and Sonnet still costs more on every step. On RAG it is $27.31 against $11.80 per 1,000 tasks. Two things drive that:
- Tokenizer. Anthropic's tokenizer produced 1.39 times as many tokens as OpenAI's for the same English document.
- Token counts on the task. On the RAG task, Sonnet used 5,184 input tokens against 3,690 for Sol, and 1,079 output tokens against 442. Sonnet also logged 615 reasoning tokens on that run.
Check the token counts for your own prompts, not only the price sheet.
What this means for choosing a model
Classification is cheap. Even Claude Fable 5.1 classifies 100 tickets for 4.5¢ in my numbers. Quality and reliability should decide this step, not price.
RAG is the cost center. For every model in the table, the RAG answer is the most expensive step per ticket. For Claude Sonnet 5.5 it is $27.31 per 1,000 tasks, against $3.97 for an email reply. The cost is driven by input size, so sending fewer or shorter documents makes every model cheaper. Retrieval quality matters as much as the model.
Email replies sit in the middle. They are short, so they stay cheap on small models. The step still grows quickly on premium ones, reaching $19.87 per 1,000 on Fable.
Reasoning tokens can surprise you. A cheap model that reasons at length can cost more than its list price implies. Look at the reasoning count, not just input and output.
Hosting, retrieval infrastructure, human review and engineering time are not included in any of these figures.
Run your own tickets through two or three candidates before committing. The cost calculator lets you plug in your own token counts and volume. Per-task detail is on the RAG answer page, the ticket classification page, the email reply page and the chatbot answer page. For model specifics, see GPT-6 Luna and Claude Sonnet 5.5. I cover price changes in news, and consumer AI subscriptions are compared on the plans page.
FAQ
How much does an AI support bot cost per ticket?
It depends on the model and the step. A RAG answer costs $0.54 per 1,000 tasks on GPT-6 Luna and $136.54 on Claude Fable 5.1. Token prices, tokenizer differences and reasoning tokens explain most of the spread.
Which step costs the most in a support bot?
The RAG answer, because it sends five retrieved documents as input. For Claude Sonnet 5.5 it is $27.31 per 1,000 tasks, against $3.97 for an email reply. Classification costs $8.90 per 1,000 runs of 100 tickets each.
Can I cut the cost without changing models?
Yes. Send fewer or shorter retrieved documents, and route simple tickets away from the retrieval step. Both reduce the input tokens that make up the largest share of the bill.