How do I calculate the cost of an AI feature?

Measure cost per completed task, not per token. Run the feature 30-50 times on realistic inputs, log input, output and cached tokens per run, multiply by current rates and take the median and the 95th percentile. The 95th percentile is the number that matters, because heavy users live there and they determine whether your pricing survives. A reference 30K-input/5K-output task costs about $0.041 on Gemini 3.7 Flash and about $0.275 on Claude Opus 5 as of August 2026.

What gross margin should an AI product target in 2026?

Aim for 70% or better on the AI cost line at your 95th-percentile user, not your median. Traditional SaaS targets 80%+, and AI features usually cannot reach that early because inference is a genuine variable cost. Below 50% at the 95th percentile you are effectively subsidizing your heaviest customers, which is the exact segment most likely to grow. Model prices have fallen sharply and repeatedly, so margins tend to improve with time if your architecture lets you switch models.

Should I pass model price cuts on to customers?

Not automatically, and not by repricing your product. Price cuts in 2026 were frequent and large — GPT-5.6 Luna dropped 80% on July 30, 2026 — but they also reverse: Gemini 3.7 Flash's $0.75/$3.75 is an introductory rate through December 31, 2026, doubling to $1.50/$7.50 on January 1, 2027. Absorb volatility inside your credit conversion rate and let improved margin fund more generous limits rather than a price change you may have to undo.

How do I stop a single customer from destroying my margin?

Enforce limits before the request, not after the invoice. Check the customer's remaining balance synchronously at the call site, cap concurrency per tenant, put a hard ceiling on agent run length and retry counts, and cap the context you will send regardless of what the user pastes. Most catastrophic AI bills come from unbounded loops and unbounded context, not from many normal users.

Quick Answer

How to Price an AI Product Around Token Costs (2026)

Published: August 24, 2026

The Short Answer

Pricing an AI product is a measurement problem before it is a strategy problem. Do it in this order:

Measure cost per completed task at median and 95th percentile
Set a margin floor against the 95th percentile, not the median
Design a credit unit that hides provider pricing from customers
Instrument enforcement at the call site
Re-verify quarterly, because model prices move constantly

Step 1: Measure Cost Per Completed Task

Per-token prices are not a cost model. A model that thinks longer can be cheaper per token and more expensive per task — measured cost per completed task on one August 2026 agentic benchmark set came out at $0.84 for Grok 4.6, $1.23 for GPT-5.6 Sol and $2.34 for Claude Opus 5, an ordering you cannot derive from list prices alone.

Do this:

Run the feature 30–50 times on realistic inputs, not toy prompts
Log input, output and cached tokens separately for each run
Multiply by current rates; compute median and p95
Include failed runs and retries — they consume tokens and they happen

Reference costs, 30K input / 5K output, verified August 2026:

Model	Input / Output per MTok	Cost per task
GPT-5.6 Luna	$0.20 / $1.20	$0.012
Gemini 3.7 Flash*	$0.75 / $3.75	$0.041
GLM-5.3	$1.40 / $4.40	$0.064
Grok 4.6	$2 / $6	$0.090
Claude Sonnet 5	$2 / $10	$0.110
GPT-5.6 Terra	$2 / $12	$0.120
Claude Opus 5	$5 / $25	$0.275
GPT-5.6 Sol	$5 / $30	$0.300

*Introductory rate through December 31, 2026; $1.50/$7.50 from January 1, 2027.

The 25x spread across that table is the entire reason pricing must be decoupled from any one model.

Step 2: Set the Margin Floor at p95

The median user is not your risk. Your risk is the customer who automates.

Target ≥70% gross margin on the AI cost line at the 95th percentile. If p95 cost per month is $18 and you charge $29, your floor holds at 38% — too thin. Either raise the price, cap the usage, or route the heavy path to a cheaper model.

Traditional SaaS margins of 80%+ are reachable over time, because inference prices have fallen repeatedly and hard, but planning your business on a future price cut is planning on someone else’s roadmap.

Step 3: Design the Credit Unit

Never expose provider tokens to customers. Define your own unit and publish a conversion.

Rules that hold up:

One credit ≈ one meaningful user outcome (a document analyzed, an agent run, a review completed) — not a token count
Price credits above your p95 cost, so model choice is yours to optimize
Keep the conversion rate internal and adjustable, so a provider change is a config edit rather than a pricing announcement
Publish typical consumption, not the formula: “most teams use 200–400 credits/month” is forecastable; a token table is not

This is what buys you the freedom to route a request to Gemini 3.7 Flash today and GPT-5.6 Luna tomorrow without touching your price list.

Step 4: Enforce at the Call Site

Every catastrophic AI bill has the same shape: a loop with no ceiling, discovered on an invoice.

Guardrails, in priority order:

Synchronous balance check before the request. After-the-fact metering is accounting, not enforcement.
Hard cap on agent iterations. Every autonomous loop needs a maximum step count that fails closed.
Retry budget per task. Retries are unmetered spend in most implementations.
Context ceiling. Truncate or reject oversized inputs regardless of what a user pastes; a 1M-token context window is an invitation, not an instruction.
Per-tenant concurrency limit. Stops one customer consuming your rate limits and your margin simultaneously.
Alert on p99 deviation, not on monthly totals. By month end the money is spent.

Step 5: Re-Verify Quarterly

Model pricing in 2026 changed constantly and in both directions:

July 30, 2026 — GPT-5.6 Luna cut 80% to $0.20/$1.20; Terra cut 20% to $2/$12
August 13, 2026 — Gemini 3.7 Flash launched at half of 3.6 Flash’s rate, introductory through year end
August 16, 2026 — DeepSeek V4 moved to peak/off-peak pricing, with peak at 01:00–04:00 and 06:00–10:00 UTC and all other hours at half that
~August 10, 2026 — Anthropic confirmed Claude Sonnet 5’s $2/$10 is permanent, cancelling a scheduled September 1 increase

A quarterly re-measurement of cost per task on your top three features takes an afternoon and routinely finds double-digit percentage savings from routing changes alone.

The Three Mistakes

Pricing off the median. Your p95 users are your growth accounts, and they are the ones who break the model.

Hard-coding provider pricing into customer-facing terms. Every price change becomes a contract change.

Optimizing before measuring. Teams routinely migrate to a cheaper model and discover cost went up because the cheaper model needed more attempts. Measure cost per completed task, then decide.