How AI Stack Cost Calculator works

1. Scope

The AI Stack Cost Calculator estimates the fully-loaded monthly infrastructure cost of an AI-powered application at four user scales (100, 1K, 10K, 100K). It prices six categories — hosting, database, auth, AI inference, email, monitoring — plus domain and any custom line items. It does not model negotiated enterprise rates, committed-use discounts, prompt caching, or usage-spike smoothing. It is a snapshot, not a live pricing feed.

2. Inputs and outputs

Inputs: hosting provider, database, auth, AI model (with average input/output tokens and API calls per user per day), email, monitoring, annual domain cost, and any custom monthly line items. Outputs: per-user monthly cost and total monthly cost at each of four user scales, plus the dominant-cost-driver insight at the 10K tier.

Engine source: src/lib/ai-stack-cost-calculator/engine.ts. Catalogs for provider tiers (HOSTING_OPTIONS, DATABASE_OPTIONS, AUTH_OPTIONS, AI_MODEL_OPTIONS, EMAIL_OPTIONS, MONITORING_OPTIONS) live in the same file and carry the as-of-date on each refresh.

3. Formula / scoring logic

# AI inference cost (the usual dominant driver)
calls_per_month = api_calls_per_user_per_day * users * 30
input_cost      = avg_input_tokens  * input_price_per_million  / 1_000_000
output_cost     = avg_output_tokens * output_price_per_million / 1_000_000
ai_monthly      = (input_cost + output_cost) * calls_per_month

# Usage-scaled hosting
hosting_monthly = base_cost + max(0, users - included_users) * per_user_cost

# Auth: MAU-scaled above the free threshold
auth_monthly    = base_cost + max(0, users - free_threshold) * per_mau_cost

total = hosting + database + auth + ai_monthly + email + monitoring + domain/12 + other

4. Assumptions

• 8 API calls per user per day is the default. This reflects a typical chat-style product with 1–3 sessions per day and some background calls. Heavy agentic products (10× higher) and notification-only products (10× lower) require overriding this input.
• Token counts are user-entered point estimates. There is no internal distribution; a product whose prompt size varies widely across calls should enter a usage-weighted average.
• Pricing is list-price only. Anthropic, OpenAI, and Google publish enterprise and committed-use discounts that can cut inference cost 20–50%. The tool does not apply these.
• Scale tiers are linear — no step-function jumps for dedicated instances, reserved capacity, or on-prem deployment.
• Auth is MAU-priced above the free threshold. Clerk Free covers 10K MAU; Auth0 Free covers 7.5K MAU; Supabase Auth is included in the database plan.
• Hosting cost scales with users for usage-based providers (Railway, Fly.io) and is flat for plan-based providers (Vercel, Render, DigitalOcean).

5. Data sources

All pricing is sourced from vendor pricing pages, dated 2026-04-24:

• Anthropic: anthropic.com/pricing
• OpenAI: openai.com/api/pricing
• Google Gemini: ai.google.dev/pricing
• Vercel: vercel.com/pricing
• Supabase: supabase.com/pricing
• Clerk: clerk.com/pricing
• Resend: resend.com/pricing
• PlanetScale: planetscale.com/pricing
• Neon: neon.tech/pricing
• Fly.io: fly.io/docs/about/pricing

6. Known limitations

• Stale pricing. The tool carries an AS_OF_DATE constant. When the date is more than 90 days old, the tool surfaces a refresh warning — this is the primary failure mode for an API-pricing tool in a fast-moving market.
• No negotiated-rate modelling. Anthropic Scale, OpenAI Enterprise, and committed-use discounts on hyperscalers can cut inference cost 20–50%. The tool takes list prices at face value; users on enterprise plans should override with their negotiated rates.
• No prompt-caching or context-window-reuse modelling. For products that reuse a large system prompt across many calls, Anthropic's prompt caching can cut input tokens by 80%+. Reflect this manually by reducing the avgInputTokens figure.
• No usage-spike smoothing. The tool assumes steady-state per-user usage at each scale. Viral growth spikes, cron-triggered batch workloads, and regional bursts will produce bills the tool does not anticipate.
• Per-user hosting cost is approximate. Vercel Pro is flat $20/mo up to bandwidth and function-execution quotas — the tool does not model those secondary caps.
• No cost for dev-time services (CI/CD, feature flags, analytics vendors). Add these as custom line items if they are non-trivial.

7. Reproducibility

8. Change log

• 2026-04-24methodology page first published. Pricing snapshot 2026-04-24 across Anthropic, OpenAI, Google, Vercel, Supabase, Clerk, Resend, PlanetScale, Neon, Fly.io.