How should an AI startup price when inference costs keep moving?

1. Blog
2. Conversion & Pricing

By Saroj Jha · November 2, 2026 · 14 min read

An AI startup should price on the value of the work its product does, then check that price against a cost floor calculated at the model prices you will pay next quarter, not the ones you pay today. Model prices for a given level of capability have fallen fast on average^[1], but individual prices can also rise on a fixed date, as three Google models are scheduled to do in December 2026 and January 2027^[2]. So build the price from cost per completed task, keep a usage allowance and an overage rate in the plan so heavy use pays its own way, size the free tier at the higher cost, and treat a model price cut as margin to bank or reinvest, not as a reason to cut your list price straight away.

The sections below take the decision in order: the floor, the margin, the effect of price changes, the model, the free tier and what to do when prices move again.

Running case: Example Co. Fictional

Example Co. is a fictional Seed-stage startup whose AI assistant triages and drafts replies to every inbound customer email for e-commerce support teams. It sells five seats at $60 each, so a typical account pays $300 a month with unlimited emails. A typical account sends 8,000 emails a month through the assistant and a heavy one 20,000. Each email uses about 6,000 input tokens and 800 output tokens, 80% of emails go to a small model and 20% to a larger one, and a quarter of drafts are regenerated before an agent accepts them.

What is your cost floor, and how do you calculate it?

Your cost floor is the inference cost of one task the customer actually accepts, including the retries it took to get there. Cost per call understates it. ICONIQ’s 2026 report includes one builder who saw “a workflow budgeted at ten cents a run drift past a dollar fifty as agents retried” and self-corrected^[3]. AAJ’s AI Product Unit Economics 2026 report walks through caching, batching and retries in detail, so here I only use the result: price from cost per completed task.

To calculate it, multiply tokens per task by the provider’s per-token prices, weight by how your traffic is routed across models, and multiply by your retry rate. Use prices read from the provider’s own page and write down the date you read them, because they change. On 1 November 2026, Google listed Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens, and Gemini 4 Argon at $2.00 and $10.00^[2]. On 2 November 2026, OpenAI’s API pricing page listed gpt-6-astra at $10.00 per million input tokens and $50.00 per million output tokens and gpt-6-luna at $0.10 and $0.50^[4], and Anthropic’s pricing page listed Claude Sonnet 5.5 at $2 and $10 and Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens^[5].

Example Co., step 1 Fictional

(Fictional usage, Google prices read 1 November 2026.) One email on Gemini 3.8 Flash costs 6,000 ÷ 1M × $0.75 + 800 ÷ 1M × $3.75 = $0.0075. On Gemini 4 Argon it costs $0.0200. With 80% routed to Flash, the blend is 0.8 × $0.0075 + 0.2 × $0.0200 = $0.0100, and with 25% of drafts regenerated, cost per completed email is $0.0100 × 1.25 = $0.0125.

Do this

Log tokens, model and retries per task for two weeks of real traffic, and calculate cost per completed task from that log, not from a test prompt.

What gross margin should you price for?

Reported AI product margins are lower than the software margins most investors are used to. ICONIQ’s Q2 2026 survey of about 300 executives at software companies building AI products found gross margins of 45% in 2025, with respondents projecting 53% for 2026 and 59% for 2027; only the 2025 figure is a reported result^[3]. Two-thirds reported better per-query unit economics, helped by managing inference costs and routing between models^[3].

Public company filings show the same pressure on a single company’s numbers. Duolingo reported that its fourth-quarter 2024 gross margin fell about 120 basis points year on year to 71.9%, in part from “increased generative AI costs due to the increased adoption of Duolingo Max”, its highest subscription tier^[6]. Duolingo is one consumer company, not a benchmark, but it shows that an AI feature priced as a premium tier still carries its cost into the margin as adoption grows.

My rule of thumb, not a benchmark: set a target margin after inference for your heaviest realistic customer, not your average one, because the average hides the accounts that lose money. A flat price with unlimited use makes that heavy customer your margin problem.

Example Co., step 2 Fictional

On the $300 seat plan, the typical account’s 8,000 emails cost 8,000 × $0.0125 = $100, a margin after inference of 66.7%. The heavy account’s 20,000 emails cost $250, leaving 16.7%. The average looks fine; the heavy account barely covers its inference before hosting, support and payment fees.

Do this

Calculate margin after inference for your top 10% of accounts by usage, and set your target against that group.

How do model price changes change the pricing decision?

Over time, the price of a given capability has fallen quickly. Epoch AI measured how fast the price of reaching fixed performance milestones fell across six benchmarks over three years, and found declines “ranging from 9x to 900x per year”, with the fastest drops in the most recent year and “less clear that those will persist”^[1]. That is a trend in the price of a capability, not a promise that the model you use today will get cheaper.

Individual prices can go up. On 1 November 2026, Google’s pricing page showed Gemini 4 Argon moving to $4.00 per million input tokens and $20.00 per million output tokens from 1 December 2026, and Gemini 3.8 Flash and 3.6 Flash doubling from 1 January 2027^[2]. A plan priced on today’s rates can lose half its margin on a date you can already read.

Costs that fall do reach the margin. In November 2025, Duolingo raised its full-year gross margin guidance by 20 basis points, which it said “reflects lower-than-expected AI costs”^[7]. In my view this is the right way to think about a cut: it improves the margin you planned on; it is not, by itself, a reason to reprice.

Example Co., step 3 Fictional

(Fictional usage, Google prices read 1 November 2026.) After both scheduled increases, one Flash email costs $0.0150 and one Argon email $0.0400, so cost per completed email rises to (0.8 × $0.0150 + 0.2 × $0.0400) × 1.25 = $0.0250, double today’s. On the seat plan, the typical account’s margin falls from 66.7% to 33.3%, and the heavy account goes from 16.7% to minus 66.7%.

Do this

Check each provider’s pricing page for scheduled changes each quarter and recalculate your floor at the highest price already announced.

Usage, seat or outcome: who should carry the cost risk?

AAJ’s guide to seat, usage and hybrid pricing covers the models in general, so here I look at one question: when model prices move, who absorbs the change? With pure seat pricing, you carry all of it. With usage or credit pricing, the customer carries some of it, and with credits you can change how many credits an action consumes without changing the list price. With outcome pricing, you carry the cost and the risk that the outcome is disputed.

The market has mostly chosen to share the risk. Bain’s analysis of public pricing at about 200 B2B SaaS companies in June 2026 found that “about four out of every five vendors introducing AI pricing are choosing capacity models” such as credits, and that among those adding a hybrid AI meter about 55% meter output, about 35% meter effort and only about 10% meter outcomes^[8]. ICONIQ’s survey found consumption-based pricing rising from 35% to 42% of respondents in six months and outcome-based from 18% to 23%, with companies blending 1.7 pricing models on average^[3].

Outcome pricing is attractive when your costs are falling, because you keep the gap. Bain finds it works when the outcome is observable, uniquely attributable to the AI and contractible, and that it has “found a real home in one category: customer support”^[8]. Even where those conditions hold, Bain reports that many vendors find outcome deals complex and time-consuming to negotiate and difficult to bill at scale^[8].

Example Co., step 4 Fictional

Example Co. moves to a $300 platform fee that includes 6,000 emails, plus $0.07 for each extra email. At today’s prices, the typical account pays $440 ($300 + 2,000 × $0.07) for a margin after inference of 77.3%, and the heavy account pays $1,280 for 80.5%. After the scheduled increases, those margins are 54.5% and 60.9%. The heavy account now pays for its own usage.

Do this

Write down, for your current plan, what happens to margin per account if your largest model cost doubles, and change the plan if the answer is a loss.

How big can the free tier be?

A free tier is an acquisition cost, so cap it at what your conversion can pay for. GitHub’s free Copilot tier, launched in December 2024, gave personal accounts “2000 Code Completions and 50 chat messages per month”^[9], a public example of a hard monthly cap on the expensive action. The cap belongs on the action that drives your inference cost, and it should be set at the post-increase cost, not today’s.

The arithmetic is simple. Decide how much margin from each converted customer you will spend on free users, multiply by your conversion rate to get a budget per free sign-up, and divide by the expected cost of one free user’s usage over their time on free. The AAJ report covers the operational side: keeping free traffic in a separately capped workspace or project and setting alerts below the provider’s hard cap.

Example Co., step 5 Fictional

If 4% of free users convert and Example Co. allows three months of a typical converted account’s margin after the scheduled increases, $240 a month ($440 − 8,000 × $0.0250), the budget per free sign-up is 0.04 × 3 × $240 = $28.80. Free users stay three months, half are active in a month, and active users use 60% of the cap. At $0.0125 per email, the cap can be $28.80 ÷ (3 × 0.5 × 0.6 × $0.0125) = 2,560 emails a month; at $0.0250 after the increases, 1,280, the same as at 2% conversion on today’s cost; and at 2% conversion after the increases, 640. Example Co. sets the free cap at 500 emails a month, inside every case.

Do this

Calculate your free-tier cap at the highest announced model price and at half your expected conversion rate, and set the cap below both.

What should you do when a model provider cuts or raises prices?

When a provider cuts prices, you have three choices: keep your price and bank the margin, spend the saving on a larger free tier or allowance, or cut your price. My view is that the first two are usually right at Seed, because your price should follow the value of the work to the customer, and competitors will see the same cut at the same moment. Cut your price when a competitor’s lower price is costing you deals you can name, not because your cost fell.

When a provider raises prices, or schedules a rise, move first on routing and caching, then on allowances and overage rates, and last on list prices for existing customers. AAJ’s article on raising prices on existing customers covers that last step.

Example Co., step 6 Fictional

If its blended cost fell 50% from today’s level, Example Co.’s margin after inference on the new plan would rise to 88.6% for a typical account and 90.2% for a heavy one. It keeps the list price, doubles the free cap from 500 to 1,000 emails while staying inside the 1,280 limit for 2% conversion at today’s cost, and reviews the decision when the scheduled increases arrive.

Do this

Write a one-page pricing policy that says in advance what you will do with a cost cut and with a cost increase, so the decision is not made in a hurry.

What changes by stage?

• Pre-seed — Log tokens, retries and model per task from the first user. Price on value to your design partners and check the floor monthly; a flat price is fine while volumes are tiny.
• Seed — Move to a platform fee with an included allowance and an overage rate before your first heavy account arrives. Cap the free tier at the post-increase cost and write the price-change policy.
• Series A–B — Report margin after inference as its own line, by segment and by model. Negotiate committed-use pricing with providers once volumes justify it, and recalculate the floor every quarter.

Free tool

Check your margin before model prices move

The AI Inference Cost Modeler takes your plan price, tasks per user, tokens per task, model prices and routing, and returns cost per task, margin per account at normal and spike usage, a break-even price and the usage level at which an account becomes unprofitable.

For the evidence on per-query costs, caching, batching and spend caps, read the AI Product Unit Economics 2026 report. For the free tier, the AI Free-Tier Cost Limit Calculator runs the cap arithmetic above on your own numbers, and SaaS pricing models in 2026 covers seat, usage and hybrid pricing in general.

If you want me to look at your own funnel, positioning and channels, my Growth Audit ends with three prioritized fixes and a 30-day action plan.