You're thinking about adding AI to your product — or you're already building one from scratch. The demos look great. The technology works. But somewhere between "let's do this" and signing off on the budget, a question keeps surfacing that nobody gives you a straight answer to: how much will this thing actually cost every month to run?
Not to build. Not to integrate. Just to keep the AI alive and responding to your users.
The honest answer is: far less than most people assume, and far more than the sticker price suggests — depending entirely on how you use it. This post walks through exactly what you'll pay, using the models that actually exist in 2026, with real scenarios you can map to your product.
First: how AI billing actually works
Every time your product calls an AI model — your user sends a message, requests a summary, asks for a recommendation — you're charged for tokens. A token is roughly three-quarters of a word. The AI reads your input (your user's question plus your instructions) and writes an output (the response). You pay for both, and output tokens cost more because generating text requires more computation than reading it.
Prices are quoted per million tokens, which sounds technical until you realize that a typical chatbot conversation uses about 1,200 tokens total. A million tokens is roughly 833 full conversations. The math starts making sense quickly.
The current model landscape as of early 2026 has three tiers you need to know about.
Premium models — GPT-5.2 ($1.75 input / $14 output per million tokens), Claude Sonnet 4.6 ($3 / $15), and Gemini 3.1 Pro ($2 / $12) — deliver the best reasoning, nuance, and reliability. Use these for complex tasks where quality genuinely matters.
Mid-tier models — Claude Haiku 4.5 ($1 / $5), GPT-5 mini ($0.25 / $2) — are strong everyday workhorses. They handle 70–80% of business tasks with quality that's nearly identical to premium, at a fraction of the cost.
Budget models — GPT-5 nano ($0.05 / $0.40), Grok 4.1 Fast ($0.20 / $0.50) — cost almost nothing. For simple, high-volume operations like classification, short summaries, or FAQ-style responses, these perform surprisingly well. GPT-5 nano is 97% cheaper than Claude Sonnet 4.6 for the same task volume.
The right choice almost never starts at the premium tier. More on that below.
What your monthly bill looks like: four real scenarios
Customer support chatbot (5,000 conversations/month)
This is the most common AI feature businesses add first. Your chatbot handles inbound questions — order status, account queries, product questions — and escalates the complex ones to a human.
A realistic support conversation runs about 1,200 tokens on a single exchange. But real chatbots send system instructions on every message (another 1,500–2,000 tokens), so the actual cost per conversation is closer to 3,000–4,000 tokens total.
At 5,000 conversations per month with Claude Haiku 4.5: roughly $25–$35/month. With GPT-5.2: $130–$175/month. With GPT-5 nano: under $3/month for simple FAQs.
The gap is enormous, and for the majority of support queries — "where is my order?", "how do I reset my password?", "what's your return policy?" — budget and mid-tier models perform identically to premium. Intercom's Fin AI Agent, which now handles over a million customer conversations weekly, lets businesses run on a managed platform instead — charging $0.99 per resolved conversation. At 5,000 monthly resolutions, that's $4,950/month, not counting base subscription. The economics only favour that approach when you factor in the hours saved building and maintaining your own chatbot.
The decision: For simple, repetitive support questions, start with GPT-5 nano or GPT-5 mini. You'll pay $3–$20/month at 5,000 conversations, and the quality difference from premium is hard to notice.
Document analysis (1,000 documents/month)
Your users upload contracts, invoices, financial statements, or intake forms. Your product reads them, extracts key information, and surfaces what matters.
A five-page document is roughly 8,000–10,000 input tokens. Generating a structured summary adds another 500–1,000 output tokens. Processing 1,000 documents monthly means about 9 million tokens.
With GPT-5 mini: roughly $330/month. With Claude Sonnet 4.6: $960/month. With GPT-5 nano: under $50/month for extraction tasks that don't require nuanced reasoning.
Allianz Insurance used AI document processing to cut claims settlement time by 80%. They didn't use the most expensive model for every document — they matched model quality to document complexity. Routine invoices went to budget processing; disputed claims with ambiguous language went to more capable models. That model routing alone typically cuts document processing costs by 40–60%.
The decision: If your documents are structured (invoices, forms, templates), start cheap. If they're complex and ambiguous (legal contracts, medical records), invest in a mid-tier or premium model for accuracy.
AI-assisted content generation (10,000 pieces/month)
You're generating product descriptions, personalised email copy, post summaries, or report drafts at scale. Each piece is roughly 2,000 input tokens (context + instructions) and 400 output tokens.
At 10,000 pieces monthly with GPT-5.2: $630/month. With GPT-5 mini: $58/month. With GPT-5 nano: under $10/month for templated content where creativity matters less than accuracy.
Shopify Magic, embedded directly into the Shopify merchant dashboard, generates product descriptions, email marketing, and image editing for millions of stores. One merchant reported generating 500 product descriptions in under an hour — tasks that previously took days. At Shopify's scale, cost optimisation matters enormously; they route the vast majority of content tasks to smaller models.
The real cost trap here isn't the per-token rate — it's the system prompt. If you're sending 3,000 tokens of instructions on every single generation request, that doubles your bill. Prompt caching (available from OpenAI and Anthropic) stores those instructions server-side and charges you 90% less for repeated use. One developer publicly shared going from $720/month to $72/month after implementing caching — a 90% reduction with no code changes beyond adding a caching flag.
The decision: For templated, structured content, budget models are more than sufficient. Spend on premium only for creative, brand-sensitive writing where tone genuinely matters.
AI-powered search (500,000 queries/month)
Search is different from the other scenarios — it usually runs on a specialised platform rather than a raw LLM API.
If you're building semantic search using vector embeddings, the main ongoing cost is the embedding model. OpenAI's text-embedding-3-small runs $0.02 per million tokens — processing 500,000 queries of 100 tokens each costs roughly $1/month. The bigger cost is vector database hosting (Pinecone, Weaviate), which starts around $25–$70/month for mid-scale use.
If you're using a managed AI search platform like Algolia, expect to pay roughly $1.75 per 1,000 search requests with AI features enabled. At 500,000 monthly queries, that's $875/month — about 3.5× what pure keyword search costs.
Zenni Optical switched to AI-powered search and saw a 34% gain in search revenue and a 27% increase in revenue per session. When AI search directly lifts conversion, the ROI math on even $875/month is easy.
The decision: If your users are already searching and not finding what they need, this is one of the highest-ROI AI additions you can make. Embedding-based search is cheap to run; managed SaaS platforms are more expensive but faster to deploy.
Where bills actually go wrong
Ignoring what's inside the request. Most developers focus on the visible conversation — what the user sends and what the AI replies. But the system prompt (your instructions to the AI), the conversation history (all previous messages in the thread), and any injected context (product data, user profile, knowledge base) all add tokens to every single request. A chatbot with a 2,000-token system prompt costs 60–70% more per conversation than the raw conversation cost alone. Always measure total tokens per request, not just the message exchange.
Starting with the most capable model. There's a natural pull toward GPT-5.2 or Claude Sonnet 4.6 because they're the best. But for most business tasks — support FAQs, product descriptions, invoice extraction — a budget model performs identically. One company processing customer feedback switched from GPT-5.2 to GPT-5 nano and saved 97% with no measurable quality difference. Always test cheaper options before committing.
No cost alerts, no spending caps. AI costs scale with usage, and usage spikes without warning. A team whose chatbot usage doubled over a holiday weekend discovered a $4,000 bill they weren't expecting. Every major API provider (OpenAI, Anthropic, AWS Bedrock, Google Vertex) lets you set hard spending limits and email alerts. Setting a soft alert at 75% of your budget and a hard cap at 110% takes ten minutes and can save thousands.
Forgetting about conversation history. Multi-turn chatbots that keep the full conversation context grow more expensive with every message. A ten-message conversation might cost 8× as much as a one-message query because every new message re-sends the entire history. Truncating or summarising older messages is standard practice, and can cut chatbot costs by 50–70% for long conversations.
A simple way to estimate your number
Start with this back-of-envelope approach. Identify your primary AI feature and its main operation. Estimate how many times it runs per month. Multiply by your realistic token count (including system prompts and context). Pick the cheapest model tier that would produce acceptable quality — then price it at that model's output rate.
For most small-to-mid-scale products — say, a chatbot handling 3,000 conversations per month, or generating 5,000 pieces of content — the monthly AI API cost lands between $10 and $200 using sensible model choices. Once you're at 50,000+ conversations, or running multiple AI features simultaneously, you start thinking in thousands per month, and cost optimisation becomes a real engineering concern.
The companies paying the most for AI aren't paying for tokens — they're paying for complexity they didn't plan for: long system prompts on every call, full conversation history on every message, premium models for tasks that didn't need them. All of it is avoidable with a bit of upfront thought.
AI running costs in 2026 are genuinely accessible. A thoughtfully built feature using the right model tier costs less per month than most SaaS subscriptions. The risk isn't the baseline expense — it's the compounding of small oversights that nobody noticed until the invoice arrived.
Start with the cheapest model that could plausibly do the job. Measure actual token counts in production. Set a spending cap before you go live. Everything else is optimisation you can do after you see real usage data.
