Every week, a founder tells us they want to "add AI" to their app. The conversation usually starts with a number they found on OpenAI's pricing page — a few dollars per million tokens — and an assumption that the hard part is already solved. It is not. API costs are real, but they are the smallest line item in a serious AI integration. Here is what actually goes into the budget.
The API Bill: Real but Misunderstood
Token costs vary dramatically by model tier, and the model you think you need is rarely the model you should use for every task. As of early 2026, the landscape looks roughly like this:
- Frontier models (GPT-5.4, Claude Opus, Gemini Ultra): $10-30 per million input tokens, $30-60 per million output tokens
- Mid-tier models (Claude Sonnet, GPT-5.4 mini, Gemini Flash): $0.50-3 per million input tokens, $1.50-10 per million output tokens
- Lightweight models (Haiku, GPT-5.4 mini, open-source hosted): $0.05-0.50 per million input tokens
Most production systems use a mix. You route simple classification tasks to a cheap model and reserve the expensive ones for complex reasoning. A well-architected system might handle 80% of requests with a sub-dollar model and escalate the rest. If you are sending every request to a frontier model, you are overpaying by 10-50x on the majority of your traffic.
For a typical B2B SaaS app with 5,000 monthly active users, API costs alone usually land between $200 and $2,000 per month, depending on usage patterns. That is manageable. The rest of the bill is where people get surprised.
Feature-by-Feature Cost Breakdown
Different AI features carry very different price tags, both in development time and ongoing costs.
Semantic search over your existing data requires generating embeddings, storing them in a vector database, and building a retrieval pipeline. Development time runs 80-160 hours. Vector database hosting (Pinecone, Weaviate, or pgvector on your existing Postgres) adds $50-500 per month depending on index size. For a corpus under a million documents, pgvector on a managed Postgres instance is often the most cost-effective option at roughly $50-100 per month with no additional vendor.
AI chatbot or assistant is the most requested feature and the most underestimated. A basic Q&A bot over your docs takes 120-200 hours to build properly. A production-grade conversational agent with tool use, memory, guardrails, and fallback handling runs 300-500 hours. Monthly inference costs for an active chatbot serving 10,000 conversations per month typically land between $500 and $3,000.
Content generation (drafting emails, summarizing reports, generating descriptions) is comparatively straightforward. Development time is 60-120 hours. Inference costs are moderate because outputs are typically short. Budget $100-800 per month at moderate scale.
Classification and routing (spam detection, ticket categorization, sentiment analysis) can often run on fine-tuned small models or even traditional ML. Development time is 40-100 hours. Inference costs are minimal — often under $50 per month — because inputs are short and you can use the cheapest models.
Build vs. Buy
For every AI feature, you face a build-versus-buy decision at multiple layers. You can use a managed AI service (like an off-the-shelf chatbot platform), integrate directly with model APIs, or self-host open-source models.
Managed platforms charge $500-5,000 per month and get you to market fast, but customization hits a ceiling. Direct API integration gives you full control at lower unit costs but requires engineering investment. Self-hosting open-source models (Llama, Mistral) eliminates per-token costs but introduces GPU infrastructure expenses — a single inference GPU instance runs $500-2,000 per month on AWS or GCP.
For most apps under 50,000 MAU, direct API integration with commercial models is the sweet spot. The engineering investment pays back in flexibility, and the per-token costs remain manageable.
The Hidden Costs Nobody Mentions
The line items that do not show up in any vendor's pricing page are often the largest.
Prompt engineering and iteration is not a one-time task. Expect to spend 40-80 hours on initial prompt development for a significant feature, then 10-20 hours per month on ongoing refinement as edge cases emerge. Prompts degrade as models update, user behavior shifts, and your product evolves.
Evaluation and testing frameworks are essential but overlooked. You need automated evaluation pipelines to measure accuracy, relevance, and safety before every deployment. Building a basic eval suite takes 60-120 hours. Without it, you are shipping blind.
Monitoring and observability for AI features requires specialized tooling. You need to track latency, token usage, error rates, hallucination rates, and user satisfaction per feature. Tools like LangSmith, Braintrust, or custom logging add $100-500 per month plus 40-80 hours of integration work.
Guardrails and safety include input validation, output filtering, PII detection, and abuse prevention. For any user-facing AI feature, budget 60-100 hours for implementation and ongoing maintenance.
Monthly Cost Projections
Here is what a realistic monthly budget looks like at different scales, excluding initial development costs:
Startup MVP (1,000 MAU, one AI feature): API costs $50-200, vector DB $0-50, monitoring $0-50. Total: $50-300 per month.
Growth-stage app (10,000 MAU, two to three AI features): API costs $500-2,000, vector DB $100-300, monitoring $100-300, ongoing prompt refinement 20 hours. Total: $1,500-5,000 per month including part-time engineering.
Enterprise application (100,000+ MAU, AI throughout the product): API costs $5,000-20,000, infrastructure $1,000-5,000, monitoring $500-1,000, dedicated AI engineering. Total: $15,000-50,000 per month.
The Real Cost Is Getting It Right
The biggest expense in AI integration is not compute — it is the engineering time to make AI features reliable, safe, and genuinely useful. A chatbot that hallucinates costs you more in lost trust than you saved by skipping evaluation. A semantic search feature that returns irrelevant results trains users to ignore it.
At Apptitude, we scope AI integrations by starting with the user problem, selecting the right model tier for the task, and building evaluation into the development process from day one. If you are planning an AI feature and want a realistic cost estimate for your specific use case, get in touch — we will walk through the numbers with you.