Most articles about custom AI model costs are written for enterprises spending $1M+. They talk about training runs on thousands of GPUs, datasets with billions of tokens, and teams of ML engineers pulling six-figure salaries. That is not you. You run a business making $500K to $10M in revenue. You want AI that actually knows your products, your policies, your industry. You do not want to fund a research lab.
Good news: you probably do not need a custom-trained model at all. Most businesses get everything they need from a smarter approach to existing models. But if you do need custom training, the costs have dropped dramatically since 2024.
We build custom AI systems at PxlPeak every week. These are the real numbers we quote to real clients, not theoretical ranges designed to justify a sales call.
The Three Approaches to Custom AI
"Custom AI" means different things depending on who you ask. A vendor selling you a $200K engagement has a different definition than we do. Here are the three real options, from simplest to most complex.
1. Prompt Engineering ($0 to $500/month)
This is not technically a custom model. It is a better way to use existing models like GPT-5.6 or Claude. You write detailed system prompts that include your brand voice, product details, pricing rules, and response guidelines. The model follows those instructions at inference time.
- Cost: $0 if you write the prompts yourself. $200 to $500 if you hire someone to do it properly.
- API costs: $20 to $200/month depending on volume. GPT-5.6 Terra runs $2 per million input tokens, $12 per million output tokens. Claude Sonnet 5 is similar at $3/$15.
- Best for: Businesses with straightforward products and fewer than 500 customer interactions per month.
- Limitation: The model has no memory of your actual documents. It can only work with what fits in the context window. Complex product catalogs, legal docs, and technical manuals will not fit.
2. RAG / Knowledge Base ($3,000 to $12,000 Setup)
Retrieval-augmented generation is the sweet spot. You keep your existing AI model (GPT-5.6, Claude, Llama) but connect it to a searchable database of your documents. When a customer asks a question, the system finds the relevant documents first, then passes them to the AI along with the question. The AI answers using your actual data.
- Setup cost: $3,000 to $12,000 depending on data volume and complexity. A 50-page knowledge base is on the low end. A 10,000-document product catalog with images and specs is on the high end.
- Ongoing cost: $200 to $800/month for vector database hosting (Pinecone, Weaviate, or Qdrant), API calls, and document sync.
- Best for: 80% of businesses. Customer support bots, internal knowledge assistants, sales agents that need product details, compliance chatbots.
Tools like LlamaIndex and LangChain make RAG setups dramatically easier than they were even a year ago. We use LlamaIndex for most of our chatbot deployments because the document ingestion pipeline is production-ready out of the box.
3. Fine-Tuning / Custom Training ($5,000 to $25,000)
Fine-tuning means taking an existing model and retraining it on your specific data. The model actually learns your patterns, terminology, and style. It becomes a different model. This is a permanent change to the model's weights, not just a prompt trick.
- Data preparation: $1,500 to $8,000. This is the part everyone underestimates. Your data needs to be cleaned, formatted into training pairs, validated, and deduplicated. Garbage in, garbage out.
- Training compute: $500 to $5,000 per run. Using LoRA (low-rank adaptation) cuts this by 80% compared to full fine-tuning. Most SMB jobs run on a single A100 GPU for 2 to 8 hours.
- Validation and testing: $1,000 to $3,000. You need human evaluators checking the outputs against ground truth. Automated metrics only catch about 60% of issues.
- Ongoing retraining: $500 to $2,000 every 3 to 6 months as your data changes.
RAG: The Sweet Spot for Most Businesses
We recommend RAG as the starting point for almost every client. Here is why: it gives you 90% of the benefit of a custom model at 30% of the cost. And if RAG is not enough, the data preparation work transfers directly to a fine-tuning project. Nothing is wasted.
What RAG Actually Costs (Itemized)
- Document ingestion and chunking: $1,000 to $3,000. Converting your PDFs, web pages, help docs, and spreadsheets into searchable vector embeddings.
- Vector database setup: $500 to $1,500. Pinecone starts at $70/month for production. Self-hosted Qdrant is free but needs a $50 to $150/month VPS.
- Retrieval pipeline: $1,000 to $4,000. Building the logic that finds the right documents, ranks them, and formats them for the AI. This includes handling edge cases like ambiguous queries and multi-topic questions.
- Integration and UI: $500 to $3,500. Connecting the RAG system to your website, CRM, or internal tools.
When RAG is enough (and it usually is):
- Your AI needs to answer questions about specific documents, products, or policies
- Accuracy matters more than style (the base model's writing style is acceptable)
- Your information changes frequently and needs live updates
- You have fewer than 50,000 documents in your knowledge base
- Response time under 3 seconds is acceptable (RAG adds 0.5 to 1.5s of retrieval latency)
When You Actually Need Fine-Tuning
Fine-tuning is the right call in four specific scenarios. If none of these apply to you, stick with RAG.
Industry-Specific Terminology
Medical practices, law firms, and engineering companies use specialized vocabulary that base models handle poorly. A dermatology clinic needs the AI to distinguish between "actinic keratosis" and "seborrheic keratosis" without hallucinating treatment protocols. Fine-tuning on your clinical notes and patient FAQs fixes this.
Brand Voice Consistency at Scale
Prompt engineering gets you 70% of the way on brand voice. Fine-tuning gets you to 95%. If you generate hundreds of customer responses per day and your brand voice is genuinely distinctive (think: Wendy's Twitter, not generic corporate), fine-tuning is worth it. For 50 responses per day, prompt engineering is fine.
Performance on Repetitive Domain Tasks
If your AI classifies support tickets, extracts data from invoices, or scores leads using your specific criteria, fine-tuning can improve accuracy from 80% to 95%+. The model learns your patterns rather than relying on general knowledge.
Cost Reduction at High Volume
A fine-tuned smaller model (Llama 3.3 8B, Mistral 7B) can match frontier-model performance on your specific task while costing 90% less per inference. If you process 100,000+ requests per month, the savings from switching to a fine-tuned small model pay for the training cost in 2 to 3 months.
The Hidden Costs Nobody Tells You About
Every AI vendor quotes you the fun part: the model training cost. Here is what they leave out.
Data Cleaning and Preparation (40% of Total Project Cost)
Your data is messier than you think. We have never onboarded a client whose data was ready to use as-is. Common issues: duplicate records, inconsistent formatting, outdated information mixed with current, PII that needs redaction, documents in 6 different formats. Budget $2,000 to $8,000 for data prep alone. It is usually 40% of the total project.
Ongoing Maintenance and Retraining
AI models do not age well. Your products change. Your policies update. Competitors launch new features your customers ask about. Plan for $500 to $2,000 every quarter for RAG updates or $1,000 to $4,000 every 6 months for fine-tuning refreshes. If you skip this, your AI starts giving wrong answers within 3 to 6 months.
Infrastructure for Local Deployment
Running models locally through tools like Ollama is attractive for data privacy. But hardware is not free. A single NVIDIA RTX 4090 (24GB VRAM) costs $1,600 and runs a 7B parameter model well. For 70B models you need an A100 ($10,000+) or rent cloud GPUs at $1.50 to $3.50/hour. Add a server, networking, cooling, and IT maintenance. Budget $8,000 to $20,000 upfront for a private AI deployment.
Quality Testing and Validation
You need humans reviewing AI outputs before going live. Plan for 40 to 80 hours of evaluation across different query types, edge cases, and adversarial inputs. At $50 to $100/hour for a qualified reviewer, that is $2,000 to $8,000. Skipping this step is how businesses end up on social media for AI fails.
Cloud API vs Local Deployment: Cost Comparison
This is the question we get in every discovery call. Here is the math at three usage levels.
Low volume (5,000 requests/month):
- Cloud API (GPT-5.6 Terra): ~$75/month
- Local (Ollama + Llama 3.3 8B on RTX 4090): ~$180/month (amortized hardware + electricity)
- Winner: Cloud. Not even close at this volume.
Medium volume (50,000 requests/month):
- Cloud API (GPT-5.6 Terra): ~$750/month
- Cloud API (Claude Sonnet 5): ~$600/month
- Local (Ollama + fine-tuned Llama 3.3 8B): ~$250/month
- Winner: Local starts winning. Break-even is around 30,000 to 40,000 requests/month, assuming you already fine-tuned the model.
High volume (200,000+ requests/month):
- Cloud API (GPT-5.6 Terra): ~$3,000/month
- Local (Ollama + fine-tuned Llama on 2x A100): ~$800/month
- Winner: Local by a mile. But only if you have the IT capacity to maintain the infrastructure. Downtime costs are real.
What PxlPeak Charges
We are transparent about our pricing because hidden costs are the number one complaint in this industry. Here is what we charge for custom AI projects.
RAG Setup: $5,000 to $12,000
- Data audit and preparation
- Vector database setup and optimization
- Retrieval pipeline with hybrid search (semantic + keyword)
- Integration with your existing chatbot or website
- Testing across 200+ query scenarios
- 2 weeks of post-launch tuning
- Documentation and handoff
Fine-Tuning Projects: $12,000 to $25,000
- Everything in the RAG package
- Training data curation (500 to 5,000+ examples)
- Model selection (we test 2 to 3 base models before committing)
- LoRA fine-tuning with hyperparameter optimization
- Human evaluation with 3 rounds of refinement
- Deployment to cloud or on-premise infrastructure
- Performance benchmarking against base model
Monthly Management: $500 to $1,500
- Knowledge base updates as your data changes
- Model performance monitoring and drift detection
- Monthly accuracy reports
- Quarterly retraining (for fine-tuned models)
- Priority support for issues
We require a 3-month minimum engagement, then month-to-month. Most clients stay for 6 to 12 months before transitioning to internal management with our documentation.
Frequently Asked Questions
What is the difference between RAG and fine-tuning?
RAG gives an existing model access to your documents at query time. The model itself does not change. Fine-tuning permanently modifies the model's internal weights using your training data. Think of RAG as giving someone a reference book. Fine-tuning is like sending them to trade school.
How long does a custom AI project take?
RAG setups take 2 to 4 weeks from kickoff to production. Fine-tuning projects take 4 to 8 weeks, with data preparation consuming most of that time. The actual model training is usually done in a day. The weeks before and after are data work and testing.
Will my data be used to train other models?
Not if you set it up correctly. OpenAI and Anthropic both offer API plans where your data is not used for training. For maximum privacy, we deploy open-source models like Llama locally through Ollama. Your data never leaves your network. This is the approach we recommend for healthcare, legal, and financial services clients who need private AI.
Can I maintain the AI system myself after setup?
Yes. We build every system with handoff in mind. RAG knowledge bases can be updated by anyone who can upload a document to a dashboard. Fine-tuned model retraining requires some technical skill, but we provide runbooks and training. About 60% of our clients manage their own systems after the first 3 months.
Can I add custom AI to my existing chatbot?
Usually, yes. If your current chatbot supports API integrations (most do), we can plug a RAG pipeline or fine-tuned model into it. The chatbot keeps its existing UI and conversation flow. The AI brain behind it gets smarter. Typical integration takes 1 to 2 weeks.
Is a custom AI model worth it for a small business?
If you handle more than 200 customer interactions per month and your product or service has any complexity beyond simple Q&A, yes. The ROI usually comes from two places: reducing support costs ($3,000 to $8,000/month in saved labor) and increasing lead conversion (10 to 25% improvement from faster, more accurate responses). A $5K RAG setup pays for itself in 1 to 2 months for most of our clients.