AI task costs · in pounds
What a website support chatbot costs to run at 10,000 chats a month
gpt-oss-120b is the cheapest of the 24 models priced, at £3.76 for 10,000 chats a month (0.038p per chat). The dearest, GPT-6 Astra, costs £568.04; the middle of the range is Llama 3.3 70B Instruct at £32.87.
How much does an AI customer support chatbot cost to run each month? The answer below covers price only, not quality.
Prices of from OpenRouter's public API price list · converted at the ECB reference rate of ($1 = £0.7538) · Urbano DX
Cost by model, cheapest first
The workload
- A chat has 4 customer messages of about 60 tokens and 4 replies of about 200 tokens.
- 2,000 tokens of instructions and help-centre text go with every message, and each message re-sends the conversation so far: 2,060, 2,320, 2,580 and 2,840 input tokens, 9,800 in all.
- The 6,960 tokens that repeat the previous request are read from the prompt cache at the model's cache rate. A model with no listed cache rate is charged its full input rate for them.
- 800 output tokens per chat. Writing to the cache, which some providers charge for, is not included.
| Model | Per chat | 10,000 chats a month | Rate used (USD) |
|---|---|---|---|
| gpt-oss-120bOpenAI | 0.038p | £3.76 | $0.037 in, $0.17 out, no cache rate per 1M tokens |
| Claude Haiku 5.5Anthropic | 0.057p | £5.68 | $0.10 in, $0.50 out, $0.01 cached per 1M tokens |
| GPT-6 LunaOpenAI | 0.057p | £5.68 | $0.10 in, $0.50 out, $0.01 cached per 1M tokens |
| DeepSeek V4.1 FlashDeepSeek | 0.061p | £6.09 | $0.02016 in, $0.891 out, $0.0054 cached per 1M tokens |
| Gemma 4 31BGoogle | 0.066p | £6.60 | $0.09 in, $0.34 out, $0.05 cached per 1M tokens |
| Qwen3.8 FlashQwen | 0.069p | £6.88 | $0.15 in, $0.47 out, $0.016 cached per 1M tokens |
| Mistral Small 4Mistral | 0.076p | £7.62 | $0.15 in, $0.60 out, $0.015 cached per 1M tokens |
| Gemini 2.5 FlashGoogle | 0.23p | £23.07 | $0.30 in, $2.50 out, $0.03 cached per 1M tokens |
| Gemini 3.5 Flash LiteGoogle | 0.23p | £23.07 | $0.30 in, $2.50 out, $0.03 cached per 1M tokens |
| GPT-4.1 MiniOpenAI | 0.23p | £23.46 | $0.40 in, $1.60 out, $0.10 cached per 1M tokens |
| Mistral Large 4Mistral | 0.31p | £30.83 | $0.68 in, $2.09 out, $0.07 cached per 1M tokens |
| Llama 3.3 70B InstructMeta | 0.33p | £32.87 | $0.59 in, $0.79 out, $0.295 cached per 1M tokens |
| DeepSeek V4 Pro 0423DeepSeek | 0.36p | £36.15 | $0.9553 in, $1.911 out, $0.07961 cached per 1M tokens |
| Gemini 3.8 FlashGoogle | 0.43p | £42.60 | $0.75 in, $3.75 out, $0.075 cached per 1M tokens |
| Claude Haiku 4.5Anthropic | 0.57p | £56.80 | $1.00 in, $5.00 out, $0.10 cached per 1M tokens |
| Muse Spark 1.3Meta | 0.6p | £60.26 | $1.25 in, $4.25 out, $0.15 cached per 1M tokens |
| Qwen3.8 Max (0902)Qwen | 0.92p | £92.11 | $2.00 in, $6.00 out, $0.25 cached per 1M tokens |
| Grok 4.7SpaceXAI | 1.1p | £105.23 | $2.00 in, $6.00 out, $0.50 cached per 1M tokens |
| Claude Sonnet 5.5Anthropic | 1.1p | £108.36 | $2.00 in, $10.00 out, $0.10 cached per 1M tokens |
| Gemini 4 ArgonGoogle | 1.1p | £108.36 | $2.00 in, $10.00 out, $0.10 cached per 1M tokens |
| GPT-6.1 SolOpenAI | 1.1p | £108.36 | $2.00 in, $10.00 out, $0.10 cached per 1M tokens |
| Claude Opus 5.5Anthropic | 2.2p | £216.73 | $4.00 in, $20.00 out, $0.20 cached per 1M tokens |
| Claude Fable 5.1Anthropic | 5.3p | £528.70 | $10.00 in, $50.00 out, $0.25 cached per 1M tokens |
| GPT-6 AstraOpenAI | 5.7p | £568.04 | $10.00 in, $50.00 out, $1.00 cached per 1M tokens |
How the numbers are worked out
- Each workload is written out on its page: tokens in and out, the share read from a prompt cache, audio minutes or images. Cost = workload × the model's listed rate, so you can check any figure with a calculator.
- Rates are OpenRouter's API prices in our snapshot of , the same ones on our model pricing pages. Buying from the developer directly can cost a little more or less. When rates move, the price change log records it.
- US dollars become pounds at the European Central Bank reference rate of : €1 = $1.1195 = £0.84385, so $1 = £0.7538. VAT, card fees and top-up fees are not included.
- About 0.75 English words make one token. Each model's tokenizer counts the same text a little differently, so treat a figure as a close estimate, not a quote.
- Models that reason before answering can add thinking tokens, billed as output. These figures count none, so a model with thinking switched on can cost more.
- This compares price only. It says nothing about which model does the job well: a cheap model that gets a tenth of the answers wrong can cost more in staff time than it saves.
Other tasks
- Summarise a 50-page PDF
- Draft replies to 1,000 customer emails
- Classify 10,000 support tickets
- Extract fields from 1,000 scanned invoices
- Transcribe 100 hours of calls
- Write 500 product descriptions
- Translate a 20,000-word website
- Generate 100 marketing images
- Answer questions over 10,000 internal documents
Want this running on your own data, with the cost measured? See the fixed-price AI pilot