How much does an AI MVP cost in 2026?

Prospects ask us this every week and they all want a quick number. The problem: “an AI MVP” can cost anywhere from US$8,000 to US$200,000, and without scoping what’s inside, the number doesn’t help.

This post is the answer we give on the first call, on screen.

The three tiers we see most

Tier 1 — Proof of concept (US$8,000–15,000)

A 3–4 week window to validate technical viability. Not production-ready. Scope:

One end-to-end flow with test data.
One model integration (Claude, GPT-4 or a hosted open-source model).
Accuracy and cost-per-request metrics.
Hand-off with repository plus a decisions document.

When it makes sense: when the executive committee needs to see something working before approving a larger budget. Not the natural step if you already have conviction on the use case.

Tier 2 — MVP in production (US$15,000–50,000)

Six to eight weeks. This is what we ship the most. Includes:

Minimal UI (web app or embedded into an existing product).
Integration with the customer’s identity, billing and analytics.
Basic guardrails (rate limiting, validation, audit trail).
Automated deploys, observability and shared on-call for the first 30 days.

A meaningful share of the cost is not the AI: it’s the plumbing. A customer support agent doesn’t cost much in models; it costs in normalized catalog data, order management integration, refund guardrails and the continuous evaluation pipeline.

Tier 3 — Enterprise / compliance (US$50,000–200,000)

Projects with compliance (HIPAA, PCI, GDPR, LOPDP) or sensitive data that can’t leave the customer’s network. The premium usually comes from:

On-premise or VPC-dedicated deploy.
DPA, audits and regulatory documentation.
Models with custom fine-tuning or self-hosted open-source LLMs.
Maintenance windows and explicit SLAs.

What’s not included in those numbers

Inference operating costs: Claude 3.5 Sonnet at the scale of a support agent runs US$200–1,500/month depending on volume. GPT-4o is more expensive; self-hosted open-source can be cheaper per request but eats engineering time.
Internal customer staff: someone has to own the project on the customer side. The non-cost is a cost.
Data: if the catalog/docs/logs are not clean, add 2–3 weeks of enrichment.

The biggest underestimation

That the plumbing around the model is most of the effort. An LLM is cheaper and more stable than three years ago. The difference between an MVP that converts and one that dies in a demo is still the same: how well wired the model is to the real system, and how measurable the guardrails are.

If you’re scoping budget for 2026, always ask for a per-week breakdown, not a single total. The total lies; the week doesn’t.