Cost Report — Website Growth Chat (MVP)

Date: July 2026
Scope: Infrastructure, third-party API, and operating costs for the first 90 days (MVP validation), with a scaling projection.
Source: Project ADRs 001, 003, 006, 007, 008 and the TRD.


1. Traffic assumptions (observed baseline)

Parameter Value
Unique site visitors/day ~100
Chat activation rate ~5%
Chat sessions/day (estimated) ~5
Chat sessions/month (estimated) ~150
Turns per conversation (average) 6–8
% of turns triggering RAG (retrieve_knowledge) ~60%

This baseline corresponds to a “corporate website” traffic profile with no significant spikes. Expected production concurrency is below 1 simultaneous session.


2. Estimated monthly cost breakdown

Component Provider Estimated MVP cost (~150 conv./month) Notes
Chat API compute Fly.io (fra, Frankfurt) ~$2 shared-cpu-1x, 256MB, with autoscale-to-zero
PostgreSQL database (pgvector + session state) Neon (eu-central-1) $0–5 Free tier covers <500 knowledge chunks; scales with usage
CDN / Edge / Rate limiting / WAF Cloudflare $0 Free tier: rate limiting rules, Bot Score, CDN
Cloud infrastructure subtotal $2–10 Reference figure from ADR-006
LLM — response generation and qualification Anthropic API (Claude Haiku 4.5) ~$1–2 See detailed calculation below
RAG query embeddings OpenAI (text-embedding-3-small) ~$0.02 Estimated in ADR-003: 10,000 query embeddings/month ≈ 100 tokens each
LLM observability (traces, tokens, evaluation) Langfuse Cloud (EU) $0 Free tier: 50,000 observations/month — sufficient for the full MVP period
Application logging + uptime monitoring Better Stack $0 Free tier: 3-day log retention; uptime monitor included
AI services and observability subtotal ~$1–2
TOTAL ESTIMATED MONTHLY COST (MVP) ~$3–12/month

3. LLM cost calculation (detail)

  • Model: Claude Haiku 4.5 — $0.25 / 1M input tokens, $1.25 / 1M output tokens.
  • Average context per turn: 4,000–10,000 tokens (system prompt + history + RAG + current turn), per the budget defined in the TRD.
  • Estimate at MVP volume:
    • 150 conversations/month × ~6 turns = ~900 turns/month
    • Input: 900 turns × ~5,000 tokens ≈ 4.5M tokens → ~$1.1/month
    • Output: 900 turns × ~400 tokens ≈ 360K tokens → ~$0.45/month
    • Total LLM ≈ $1.5–2/month at MVP volume

This figure is well below the configured soft cost ceiling (MONTHLY_COST_CAP_USD = $50) and the architecture review threshold defined in ADR-001 ($500/month, at which point prompt caching or model routing would be evaluated).


4. Costs not included in the recurring calculation (contingent)

Item Trigger condition Estimated cost
Better Stack — paid plan If log volume exceeds 3GB/month or retention >3 days is needed from $29/month
Langfuse — paid plan If 50,000 observations/month are exceeded Flat rate (no per-seat pricing); confirm at Langfuse.com
Neon — paid plan If the knowledge corpus grows beyond the free tier (>0.5GB) Variable, scales with usage
Prompt caching / model routing If LLM cost sustainably exceeds $500/month Reduces cached-token cost by ~80%
Migration to AWS Fargate + RDS + WAF If enterprise-grade infrastructure is required post-validation ~$60–90/month (ADR-006 reference)
DPA / legal compliance (Anthropic, OpenAI, Better Stack, Langfuse, Cloudflare) Before processing real visitor data in production Not quantified — legal/administrative cost, not infrastructure

5. Comparison: discarded alternative (reference)

For context, the infrastructure alternative evaluated and discarded (AWS Fargate + RDS db.t3.micro + ALB + WAF) had an entry cost of $60–90/month, versus $2–10/month for the chosen stack (Fly.io + Neon + Cloudflare). The decision prioritizes cost minimization and operational simplicity for a 1–2 engineer team during the MVP validation phase (ADR-006).


6. Executive summary

Scenario Estimated monthly cost
Current MVP (~150 conv./month) ~$3–12/month
With paid logging/observability (if free tier is exceeded) +$29–50/month
Scaling to 10x volume with no architecture changes ~$15–40/month (LLM cost scales linearly; cloud infra has ample headroom)
Migration to enterprise infrastructure (AWS) ~$60–90/month for compute/database alone

Conclusion: MVP operating cost is marginal (single-digit to low double-digit dollars per month), thanks to the combination of free tiers (Cloudflare, Langfuse, Better Stack, Neon) and a low-cost LLM model (Claude Haiku 4.5). The largest cost-deviation risk is conversation volume, not per-token unit price — any significant traffic scaling should trigger a revalidation of these figures.


Note: all figures are drawn from estimates documented in the project’s ADRs (001, 003, 006, 007, 008) and the TRD. They do not substitute for real production spend monitoring (Langfuse + Anthropic API alerts + Better Stack).