LLM Cost per Customer: AI SaaS Unit Economics
Prices verified 2026-07-14 against official provider pricing pages. Methodology and formula: how these numbers are computed.
Quick answer: a typical AI-assisted SaaS feature (200 requests/customer/month, 2-call chain, 60% cache hit rate) on Claude Sonnet 5 costs $2.15 per average customer — 7.4% of a $29 subscription. But with the usual usage skew (10% of users at 7x usage), a power user costs $9.41/month (32% of price), and the top decile consumes 44% of your entire LLM bill. Price for the tail, not the average — enter your numbers below.
Your feature's economics
Average customer cost by model (baseline scenario)
1,000 customers, 200 requests/customer/month, 2 LLM calls per request, 1,500 input / 400 output tokens per call, 60% cache hit rate, 10% power users at 7x usage, $29/month price. Real-time traffic — batch API does not apply.
| Model | Avg customer / month | Power user / month | % of $29 price (avg) | Total (1,000 customers) |
|---|---|---|---|---|
| DeepSeek DeepSeek V4 Flash | $0.079 | $0.347 | 0.3% | $79.41 |
| DeepSeek DeepSeek V4 Pro | $0.245 | $1.07 | 0.8% | $245 |
| Google Gemini 3.1 Flash-Lite | $0.309 | $1.35 | 1.1% | $309 |
| Anthropic Claude Haiku 4.5 | $1.08 | $4.71 | 3.7% | $1,076 |
| OpenAI GPT-5.6 Luna | $1.24 | $5.41 | 4.3% | $1,236 |
| Google Gemini 3.5 Flash | $1.85 | $8.11 | 6.4% | $1,854 |
| Anthropic Claude Sonnet 5 | $2.15 | $9.41 | 7.4% | $2,152 |
| Google Gemini 3.1 Pro | $2.47 | $10.82 | 8.5% | $2,472 |
| OpenAI GPT-5.6 Terra | $3.09 | $13.52 | 11% | $3,090 |
| Anthropic Claude Sonnet 4.6 | $3.23 | $14.12 | 11% | $3,228 |
| Anthropic Claude Opus 4.8 | $5.38 | $23.54 | 19% | $5,380 |
| OpenAI GPT-5.6 Sol | $6.18 | $27.04 | 21% | $6,180 |
| Anthropic Claude Fable 5 | $10.76 | $47.07 ⚠ | 37% | $10,760 |
⚠ = a power user's LLM cost alone exceeds the $29 subscription — that account loses money.
Break-even usage by price tier
Requests per month at which one customer's LLM cost (Claude Sonnet 5, same workload shape) equals their subscription — beyond this they are underwater before any other cost:
| Price tier | Break-even requests / month | ≈ per working day |
|---|---|---|
| $9/month | 836 | 38 |
| $19/month | 1,766 | 80 |
| $29/month | 2,695 | 123 |
| $49/month | 4,554 | 207 |
| $99/month | 9,201 | 418 |
Rule of thumb: keep average LLM cost under 20% of subscription price, so the power-user tail stays profitable without usage caps.
This calculator estimates — it doesn't meter. Knowing which real account consumed what requires per-customer cost attribution: tagging every LLM call with a customer ID and aggregating. Provider dashboards only show org-level totals; enterprise FinOps platforms start at hundreds per month. We're building a lightweight per-customer cost dashboard for exactly this gap — email us for early access and tell us your stack.
Sources: OpenAI pricing · Anthropic pricing · DeepSeek pricing · Gemini API pricing. Per-token prices and model caveats: migration savings calculator · methodology. Running an agent chain? Price retries with the agent cost calculator.
Frequently asked questions
How much does an AI feature cost per customer per month?
For a typical AI-assisted SaaS feature (200 requests/customer/month, 2 LLM calls per request, 1,500 input / 400 output tokens per call, 60% prompt-cache hit rate) on Claude Sonnet 5, the average customer costs $2.15/month in LLM spend — 7.4% of a $29 subscription. But averages mislead: with a typical usage skew (10% power users at 7x usage), your top decile costs $9.41/month each (32% of the price). Prices verified 2026-07-14; enter your own numbers in the calculator above.
How do I calculate LLM cost per customer without an observability stack?
Multiply per-request cost by requests per customer: cost_per_request = calls_per_request × (input_tokens × effective_input_price + output_tokens × output_price) ÷ 1M, where effective input price blends cache hits at ~10% of list price. This gives a solid planning estimate. For actual attribution you eventually need per-customer metering — tag each API call with a customer ID and aggregate — which is exactly the gap between provider dashboards (org-level totals only) and enterprise FinOps platforms.
What share of my LLM bill do power users consume?
With the common 10%-of-users-at-7x-usage skew, the top 10% of customers consume 44% of total LLM spend. The practical consequences: (1) average cost per customer understates tail risk — price and rate-limit for the tail, not the mean; (2) a flat subscription quietly loses money on heavy accounts long before the blended margin shows it; (3) usage caps or overage pricing on the top decile protect margin with minimal churn exposure.
At what usage does a $29/month AI subscription lose money?
On Claude Sonnet 5 with this workload shape (2 calls/request, 1,500 in / 400 out tokens, 60% cache), the LLM cost of one request is $0.011, so a $29 subscription breaks even at 2,695 requests/month. A $9 tier breaks even at 836 requests. Margin planning rule of thumb: keep average LLM cost under 20% of subscription price so the power-user tail stays profitable.
Should I meter and bill AI usage per customer?
Once any single customer's LLM cost approaches your gross-margin target, yes. Common patterns, in order of adoption: monthly usage caps with soft limits (simplest), credit packs, per-seat AI add-on pricing, and true usage-based billing. The prerequisite for all of them is per-customer cost attribution — knowing which account consumed what — which provider dashboards do not give you.
Which model should power a $19-49/month AI SaaS feature?
Work backwards from price × 20% target: at $29/month that is a ~$6 LLM budget per customer. With this workload shape, mid-tier models (Claude Sonnet 5 at $2.15, Gemini 3.1 Pro, GPT-5.6 Terra) fit comfortably; frontier models (GPT-5.6 Sol, Claude Fable 5) blow the budget on power users alone. Budget models (DeepSeek V4, GPT-5.6 Luna) leave the most margin but test quality on your own eval set first — the ranked table above compares all of them under identical assumptions.