How to Price AI Features: Usage, Seat, and Bundled Pricing in 2026

Every SaaS founder I talk to in 2026 is wrestling with the same question: how to price AI features without either torching gross margins or leaving millions on the table. The rules that worked for traditional software simply do not translate. When your marginal cost per user was effectively zero, you could charge a flat seat fee and sleep well. But AI features have real, variable, and sometimes shockingly volatile underlying costs tied to token consumption, GPU time, and third party inference providers.

The uncomfortable truth is that most early AI pricing experiments from 2023 and 2024 quietly failed. Companies launched "AI add ons" at 10 or 20 dollars per seat, discovered that 5 percent of their users consumed 80 percent of the tokens, and watched their gross margin on those customers drop below 40 percent. Some went negative. The survivors learned a hard lesson: AI pricing is not a marketing decision, it is a unit economics decision that happens to have a marketing wrapper.

In this guide I will walk through every major pricing model teams are using in 2026, the specific companies that pioneered each approach, and the exact tradeoffs I see when we advise clients on rolling out new AI capabilities. No theory, no hand waving. If you are launching an AI feature this quarter, you should leave with a clear framework for choosing a model that protects your margin while still feeling fair to customers.

Strip away the buzzwords and there are really only five pricing models being used for AI features at scale in 2026. Every hybrid you see in the wild is some combination of these primitives.

• Pure usage based. Customers pay per token, per API call, per generated image, or per minute of audio. OpenAI and Anthropic price their APIs this way, and developer tools like Cursor and v0 have adopted it for their higher tiers.
• Seat based with unlimited AI. A flat per seat fee includes all AI functionality with soft or hard fair use caps. Linear took this approach with Linear Agents, charging a flat uplift per seat regardless of how many agent runs a user triggers.
• Credit or token bundles. Customers buy a pool of credits that get consumed by AI actions at varying rates. ClickUp AI, Jasper, and Copy.ai all ship variations of this. It is the most flexible model but also the most confusing for buyers.
• Outcome based. You charge only when the AI does something measurable and valuable. Intercom Fin is the canonical example, charging roughly 0.99 dollars per resolved conversation rather than per message or per seat.
• Bundled into existing tiers. AI is included "free" in your Pro or Business plan as a differentiator, funded by the price increase that came with the plan itself. Notion did this most famously when it rolled Notion AI into its Business plan at a higher price point in 2024.

None of these is universally correct. The right choice depends on three variables: the predictability of your per customer inference cost, the perceived value of the AI output, and the sophistication of your buyer. I will come back to each of these throughout the rest of this piece.

Usage based pricing is the most defensible model when your costs are volatile and your customers are technical enough to understand tokens, calls, or compute units. It is also the riskiest from a customer psychology standpoint because buyers hate unpredictable bills.

The critical distinction nobody talks about is pass through pricing versus value pricing. Pass through means you charge customers roughly what the underlying model costs you plus a fixed margin, often 20 to 40 percent. Value pricing means you charge based on what the output is worth, regardless of the token count under the hood. A 500 token summary that saves an analyst two hours of work is worth far more than 500 tokens of chit chat.

My strong opinion: pure pass through pricing is a trap for product companies. It turns you into a reseller of OpenAI or Anthropic, it caps your margin at whatever spread you can defend, and it makes your pricing hostage to your vendors. When GPT-4 Turbo dropped in price by 60 percent in 2024, companies on pass through pricing had to decide whether to cut their own prices or risk looking greedy. Companies on value pricing simply kept the windfall.

If you do go usage based, use a proper metering infrastructure. Building this yourself is a waste of engineering time. Tools like Metronome and Orb handle the complicated parts: real time aggregation, proration, credit balances, invoice generation, and the edge cases around retries and failed inference calls. Stripe Billing has caught up on the basics but still lags on complex credit models. Expect to pay roughly 0.5 to 1.5 percent of metered revenue for a dedicated metering provider. It is almost always worth it. For a broader primer on rolling out metered billing, see our guide to usage based pricing implementation.

Credits are the dominant pricing pattern for consumer and prosumer AI tools because they solve a very human problem: buyers want to know what they are spending before they spend it. A 20 dollar credit pack is easier to stomach than an open ended bill based on "tokens."

ClickUp AI sells credits that get consumed at different rates depending on the action. A quick summary might cost one credit while a long document generation costs ten. Copy.ai and Jasper built entire businesses on word or credit packs before transitioning to hybrid models. The genius of credits is that they create breakage revenue, meaning the credits customers buy but never use, similar to how gift cards work in retail. Well designed credit systems see 15 to 30 percent breakage, which is pure margin.

The downside is that credits create friction at exactly the wrong moment. When a user hits zero and is told to buy more before finishing their workflow, you risk turning a delighted customer into a frustrated one. The fix is auto top up with clear notifications and generous rollover policies. Do not be the company that expires unused credits at the end of the month. That saves you pennies and costs you churn.

If you go with credits, be ruthless about unit transparency. Tell users exactly how many credits each action will cost before they commit. Writer, the enterprise AI platform, shows a predicted credit burn for every generation above a certain size. This small UX decision dramatically reduces support tickets and refund requests.

Seat based AI pricing is the most buyer friendly model and, if you can make the math work, often the most profitable. The principle is simple: charge a flat per seat uplift for access to AI features and absorb the usage variance across your customer base.

Notion AI launched in 2023 as a 10 dollar per member add on, then in 2024 bundled it into the Notion Business plan at a higher price point. The effect was dramatic. Attach rates went from roughly 6 percent as an add on to effectively 100 percent of Business customers, and Notion used AI as the primary justification for the price increase. Whether or not heavy users were profitable individually mattered less because the bundle moved the average revenue per customer up meaningfully.

Linear took a different but equally disciplined approach with Linear Agents. Rather than metering agent runs, Linear charges a flat uplift per seat and publishes soft fair use guidelines. The bet is that the 80 percent of users who run two or three agents per week cross subsidize the 5 percent of power users who run fifty. Jira Atlassian Intelligence uses nearly the same logic, bundling AI capabilities into their Premium and Enterprise tiers with no explicit per seat surcharge at all.

The math only works if three conditions hold. First, your inference costs must be bounded and predictable per seat, ideally under 2 to 4 dollars per user per month on average. Second, your plan pricing must leave enough headroom to absorb the top decile of usage. Third, you need contractual protection against obvious abuse, which I will cover in the next section. If you cannot meet all three conditions, seat based is how you go bankrupt slowly.

This is the section most pricing posts skip and where I see founders lose the most money. AI gross margin is not accounting gross margin. Your CFO may report 80 percent gross margin on your AI line item because that is what the software P and L looks like. The reality, once you attribute true cost of goods sold including inference, vector storage, embedding refreshes, and the fully loaded cost of the infrastructure engineers keeping everything up, is often 50 to 65 percent. On heavy users it can be negative.

You must instrument gross margin per customer, not just per product. Look at the bottom decile of profitability. In almost every AI SaaS business I audit, between 2 and 8 percent of customers are consuming enough tokens to be unprofitable. Sometimes dramatically so. The fix is not always to raise prices. Sometimes it is to add a fair use ceiling, throttle the 99th percentile, or move those specific customers to a metered plan.

Protect yourself with explicit terms of service language around automated use, bulk operations, and scripted abuse. I have seen companies lose thousands per month to a single customer running a scraping script through a chat endpoint. You also need technical guardrails: per user rate limits, per workspace caps, anomaly detection on token spikes, and the ability to throttle a specific customer without a code deploy.

Finally, bake hedging into your model choices. Route easy queries to cheaper models like Haiku, GPT-4o mini, or Gemini Flash. Reserve the expensive frontier models for queries where accuracy actually matters. Good AI product teams are now spending as much engineering effort on model routing as on prompt engineering, because the cost delta between a cheap and expensive model is often 20x or more for essentially equivalent quality on simple tasks. This kind of cost engineering feeds directly into how to calculate AI ROI in a way that executives actually believe.

Feature gating is where AI pricing meets product strategy. You have to decide which AI capabilities are free, which are gated behind a paid tier, and which are pure add ons. Get this wrong and you either give away the core value or bury it so deep that nobody finds it.

My rule of thumb: give away the demo, charge for the workflow. Let free users experience the magic moment, whether that is drafting a document, summarizing a thread, or answering a question. Gate the bulk operations, the custom models, the advanced reasoning, and the higher volume consumption. This mirrors how Intercom Fin handles trials, giving new customers a pool of free resolutions before the 0.99 per resolution meter kicks in.

For free trials specifically, cap the trial by credits or actions rather than by time. A 14 day trial of an AI feature is useless if the user only opens your product twice during that window. A trial that includes, say, 500 credits or 20 AI generations gives every user a fair shot at the aha moment regardless of when they log in.

Hybrid models are now the default for most serious AI SaaS products. A typical 2026 hybrid looks like this: a flat platform fee per seat that includes a generous baseline of AI usage, plus metered overages above the baseline, plus optional credit top ups for bursty workloads. This gives you the predictability of seat pricing for 90 percent of customers and the margin protection of metered pricing for the 10 percent who actually would break your unit economics. Writer and ClickUp have both landed on variations of this structure after experimenting with pure models and finding them lacking. Before finalizing any hybrid, sanity check it against the fundamentals in our guide on how to price a SaaS product.

Here is the decision framework I use with clients. Answer these questions in order and the right model usually becomes obvious.

• How predictable is your per customer inference cost? If the standard deviation across customers is low, seat based works. If it is high, you need metering or credits to cap downside.
• How sophisticated is your buyer? Developers and technical buyers tolerate usage based pricing. Marketing, HR, and ops buyers strongly prefer flat predictable fees.
• Is the AI the core product or a feature within a larger product? Core AI products like Writer or Jasper lean toward credits and usage. AI as a feature inside a broader tool, like Notion or Linear, leans toward seat bundles.
• What is your gross margin ceiling? If your AI costs exceed 35 percent of revenue at the median customer, you cannot afford seat based without a price increase. You will need metering to protect margin.
• How clearly can you define a valuable outcome? If the AI produces a clean, measurable result like a resolved ticket, a qualified lead, or a completed task, outcome based pricing is strictly superior to everything else. Intercom Fin proved this. Almost nobody else has pulled it off yet.

The most common mistake I see is founders choosing a pricing model based on what their competitors do rather than what their unit economics support. Your competitor may be bundling AI into their Pro plan because they raised a 200 million dollar Series C and can afford to subsidize losses for two years. You probably cannot. Price for your business, not theirs.

One final point that is easy to forget: AI pricing is not a one time decision. Model costs are still falling roughly 4x per year on a capability adjusted basis, which means the economics of whatever you ship today will look different in six months. Build pricing infrastructure that lets you experiment. Run cohort analyses every quarter. Be willing to migrate customers to new plans as the underlying cost curve shifts. The teams that will win the next three years of AI monetization are not the ones with the cleverest launch pricing. They are the ones who treat pricing as a living system and iterate on it faster than anyone else.

If you are trying to figure out how to price AI features in your specific business and want a second set of eyes on the unit economics, positioning, and rollout plan, we help SaaS teams work through this every week. Book a free strategy call and we will walk through your numbers together.