How do I protect margins when model prices change?

Price on a unit you control, not on tokens, and keep a margin floor at the 95th percentile of cost per task. Model prices move both ways: Gemini 4 Argon's introductory $2/$10 per million tokens rises to $4/$20 on December 1, 2026, and Gemini 3.8 Flash doubles to $1.50/$7.50 on January 1, 2027. Re-run your cost per task whenever a provider changes prices, and route routine work to cheaper models.

Should AI products charge per seat or per usage?

Use seats for predictable everyday features and usage for the expensive ones. Microsoft and Atlassian both do this in 2026: everyday chat and summaries are included, while agentic work (Copilot Cowork, Code and Autopilot; Rovo Chat, Jira Coding Agent) is metered. Pure per-seat pricing breaks when a few power users run agents all day; pure usage pricing scares buyers who cannot forecast.

What is outcome-based pricing for AI products?

Charging for a completed result instead of the work behind it. QA Wolf bills one run per verified pass or fail result, with up to three attempts and AI repair included; Salesforce charges $2 per Agentforce conversation; Atlassian meters AI agent resolutions where an agent completes a request without human escalation. It shifts model-cost risk to the vendor, so it only works when you can measure cost per outcome precisely.

Quick Answer

AI Product Pricing Strategies for Unpredictable Token Costs

Q: What are the best AI product pricing strategies for unpredictable LLM token costs?

Five models are in use as of November 2026. Included allowance plus metered overage (Atlassian Rovo: pooled credits, then $0.01 per credit from December 3, 2026). Prepaid balance with a hard stop (Momentic: a monthly dollar balance, runs pause at $0, no overages). Price per outcome (QA Wolf at $1 per verified web test run; Salesforce Agentforce at $2 per conversation). Seat plus usage for heavy features (Microsoft 365 Copilot seat, with Copilot Credits for advanced experiences). Bring your own key (Mem0's free plan rises to 50,000 adds a month with your own LLM key). Pick the one that moves cost variance to whoever controls it.

Published: November 5, 2026

The short answer

Do not pass raw token prices to customers. Pick a billing unit whose cost you can predict, then put the variance on whoever controls it. In 2026, five models dominate, and every large vendor uses one or two of them:

Strategy	Who carries the variance	Real example (Nov 5, 2026)
1. Included allowance + metered overage	Customer, above the allowance	Atlassian Rovo: pooled credits per plan, then $0.01 per credit
2. Prepaid balance, hard stop	Customer, capped	Momentic: monthly dollar balance; runs pause at $0
3. Price per outcome	Vendor	QA Wolf $1 per verified web run; Agentforce $2 per conversation
4. Seat + usage for heavy features	Split	Microsoft 365 Copilot seat; Copilot Credits for Cowork, Code, Autopilot
5. Bring your own key	Customer’s model bill	Mem0 Free: 50,000 adds/month with your own LLM key

The method for measuring cost per task, margin floors and credit design is in how to price an AI product around token costs. This page covers which model to choose.

Why token costs are unpredictable

Three sources, all getting larger:

Per-task variance. Agentic and reasoning workloads vary by an order of magnitude between easy and hard requests; the same prompt can emit very different numbers of hidden reasoning tokens.
Price changes you do not control. Gemini 4 Argon is $2/$10 per million tokens until November 30, 2026 and $4/$20 from December 1; Gemini 3.8 Flash goes from $0.75/$3.75 to $1.50/$7.50 on January 1, 2027. See current API prices.
Power users. A few customers running agents all day can cost more than the rest combined.

1. Included allowance plus metered overage

Bundle a generous allowance into the plan, meter beyond it. Atlassian Rovo gives 25, 70 or 150 credits per user a month on Jira and Confluence Standard, Premium and Enterprise, pooled across the organization so light users subsidise heavy ones. From December 3, 2026, usage above the pool bills at $0.01 per credit with extra usage on by default, up to an admin cap; admins get alerts at 80% and 100%. Cheap features (in-line summaries, Rovo Search) stay free, so the meter only touches expensive work.

Use it when most customers stay inside the allowance. Watch: defaulting overage to “on” protects your margin but creates surprise bills; give admins caps and alerts from day one.

2. Prepaid balance with a hard stop

Sell a balance in dollars or credits, and stop when it runs out. Momentic includes $50 of usage a month on Free and $500 on Scale; test steps, hosted minutes, AI features and its Mo agent all draw from one balance, and runs pause at $0 instead of billing overages. Scale customers top up from $50. Chroma Cloud does the same at the infrastructure level: hitting a usage limit pauses the service.

Use it when buyers fear bill shock more than downtime: developer tools, SMBs, self-serve. Watch: a hard stop in production breaks customers’ workflows; warn early.

3. Price per outcome

Charge for the result, and absorb the cost of getting there. QA Wolf bills $1 per verified web test run ($2 on mobile and desktop), including up to three attempts, AI repair of broken tests and bug filing; a flaky retry costs the customer nothing. Salesforce Agentforce charges $2 per conversation for customer-facing agents, and Atlassian meters “AI agent resolutions” where an agent completes a request without escalation.

Use it when you can measure cost per outcome tightly and it sits well below the price. Watch: you carry all the model-price risk; build in routing to cheaper models and caching.

4. Seat plus usage for heavy features

Keep a per-user price for everyday AI and meter only the agentic work. Microsoft 365 Copilot is a per-seat licence covering chat and Copilot in Word, Excel, Outlook and Teams; Copilot Cowork, Code, Autopilot and the Work IQ APIs bill by usage in Copilot Credits. Salesforce offers the opposite trade: Agentforce add-ons at $125 per user a month give unmetered employee use, so the customer buys predictability with a higher seat price.

Use it when you sell to enterprises that budget in seats.

5. Bring your own key

Let customers connect their own model provider and charge only for your software. Mem0’s Free plan covers 10,000 adds and 1,000 retrievals a month for 30 days with model calls on Mem0, then rises to 50,000 adds and 10,000 retrievals a month at no charge once the customer connects their own LLM key.

Use it when customers are developers with their own provider contracts. Watch: you lose model-margin revenue and control over quality.

How to choose

Can the customer forecast usage? No: allowance or prepaid balance. Yes: usage or outcome.
Can you measure cost per outcome within ±20%? Yes: outcome pricing is your strongest sales story.
Selling to enterprise procurement? Seats plus metered heavy features.
Whichever you pick: cap your own spend at the provider; see how to set hard spend limits on LLM APIs.

Last verified: November 5, 2026.