How to Price Your AI SaaS in 2026

Pricing an AI SaaS in 2026 means picking from four models: per-seat, usage-based, outcome-based, or hybrid. The right pick depends on what your AI produces and how variable your AI cost of goods sold is per user. This post gives a decision tree, decodes 8 real pricing pages, and shows the margin math.

This is a SoGood post and SoGood is one of the products mentioned. We are not a pricing tool. The Finance department inside SoGood can advise on pricing as part of a business case, but you would not subscribe to SoGood to price your own product. We're writing this because we did the work for ourselves and want to save you the false starts.

Why AI SaaS pricing broke in 2024

Traditional SaaS pricing assumed near-zero marginal cost per user. Storage and compute were cheap; a $20 per month seat funded itself by Tuesday and printed margin for the rest of the month. AI changed the unit economics: a single active user can burn $30 of inference per day on frontier models. The old per-seat playbook breaks the moment your power users cost more than they pay.

Three forces compress AI SaaS margin in 2026. First, variable COGS: every prompt has a real GPU bill. Second, infinite-use buyers: the highest-value users are also the highest-cost ones, so revenue and cost scale together. Third, race-to-the-bottom anchoring: ChatGPT Plus at $20 set a price ceiling that almost no AI product can clear without justifying serious extra value.

The consequence is that founders who copy the per-seat playbook end up either capping usage (and getting yelled at on Reddit), eating the margin (and going broke), or pivoting pricing within 6 months (and confusing customers). The fix is to pick the right model from the start. There are four that actually work.

The 4 pricing models that work in 2026

Model 1: Per-seat

Flat monthly fee per active user, unlimited (or soft-capped) AI usage inside. Works when your AI cost per active user stays under 20 percent of price. Linear at $16, Granola at $14, Loom Business + AI at $24 all hold this shape profitably because the AI augments a user without doing heavy compute per seat.

Pros. Predictable revenue, simple to sell, finance teams understand it, fits enterprise procurement. Cons. Breaks for compute-heavy workloads, hard to monetize power users, race-to-the-bottom pressure from ChatGPT Plus.

Model 2: Usage-based

Charge per token, per build, per generation, or per credit. Works when output is discrete and measurable: a code build, a video, an image, a brief. Bolt, v0, Cursor (on overages), and Replit all use this either as the primary model or the overage layer behind a flat subscription.

Pros. Margin scales with revenue, no power-user trap, transparent to buyers who can self-rate-limit. Cons. Revenue is bumpy, finance teams hate forecasting it, fear of a runaway bill kills trial conversions. The fix is published rate cards plus hard caps with overage opt-in.

Model 3: Outcome-based

Charge per result delivered: a closed support ticket, a published page, a qualified lead, a shipped business case. Intercom's Fin charges per resolution; SoGood charges per shipped outcome via credits.

Pros. Buyer aligns price to value, easy to justify in a procurement deck, hard for race-to-the-bottom flat-fee competitors to undercut. Cons. You have to define the outcome unambiguously, attribution disputes are common, and the model only works for products where the outcome is countable and obvious.

Model 4: Hybrid (platform fee + metered AI)

A small flat fee covers the software platform; AI usage is metered separately as credits or tokens. Lovable Pro at $25 per month with 100 credits, Notion Business at $19.50 plus a separate AI credit pack, ChatGPT Team at $25 per seat with shared message caps. This is the model most AI products converge to within 18 months.

Pros. Predictable floor revenue, variable AI cost passes through, power users self-select into bigger packs. Cons. Two pricing dimensions confuse buyers, packaging tests take longer, you need a usage dashboard from day one.

The pillar context for solo and bootstrapped founders making this call is in best AI tools for solo founders 2026; the broader shift in team shapes that makes this question urgent is covered in the 12-person startup is dead.

8 real pricing pages, decoded

These prices were verified on the public pricing pages on 2026-05-18. Pricing in this category moves quarterly. Check the source URLs before quoting.

A couple of honorable mentions for shape reference. ChatGPT Plus sits at $20 per month with fair-use caps on frontier models, subsidized by OpenAI's enterprise revenue. Claude Pro lists at $20 per month with usage caps that reset on a rolling window. Both are anchor prices, not target prices, for a bootstrapped product.

Per-tool reads (the part the table can't carry)

Cursor ($20 / $40). The cleanest example of "flat price with a usage safety valve." The $20 Individual tier captures 80 percent of solo users at a margin; power users hit overage billing on Bugbot and frontier model usage. Source: cursor.com/pricing, 2026-05-18.

Lovable ($25 / $50). The cleanest hybrid in the no-code AI builder category. 100 credits per month with a 5 daily credit ceiling, rollovers, and on-demand top-ups. The daily cap exists specifically to prevent a single user from burning a month of compute in one day. Source: lovable.dev/pricing, 2026-05-18.

Bolt ($25 Pro). A pure token meter inside a flat-fee wrapper. 10M tokens per month is roughly 30 to 50 small app builds, depending on the model and the prompt discipline. Free tier has a daily token guardrail that flat-tier Pro removes. Source: bolt.new pricing, 2026-05-18.

v0 ($30 / $100). Aggressively transparent: $30 of credits per user per month, then per-model token pricing. v0 Max Fast at $150 per 1M output tokens is the highest-priced unit on the page, which tells you exactly how much speed costs. Source: v0.app/pricing, 2026-05-18.

Linear ($10 / $16). Per-seat with AI features tier-gated. Business at $16 unlocks Linear Agent automations, Code Intelligence, and Triage Intelligence. The AI features are bundled, not metered, because they sit behind product workflows that don't burn compute per seat. Source: linear.app/pricing, 2026-05-18.

Notion AI ($19.50 + $10 per 1K credits). Notion separated AI billing from seats in 2025. The $19.50 Business seat now buys you the workspace; AI is a $10 per 1,000 monthly Notion credits add-on. The separation lets Notion price seats stably while inference costs move. Source: notion.com/pricing, 2026-05-18.

Loom AI ($24). The simplest possible hybrid: a separate $24 tier that includes AI features, sitting alongside the $18 non-AI Business tier. You buy the AI by upgrading the seat, not by metering tokens. Works because the AI surface is bounded (transcripts, summaries) and doesn't run free-form. Source: loom.com/pricing, 2026-05-18.

Replit ($25 / $100). Annual-billed flat fee with a credits-included structure. $25 of credits per month at Core, $100 at Pro, plus database rollbacks and parallel agent count as the real Pro differentiators. Source: replit.com/pricing, 2026-05-18.

The margin-killer chart

The pattern is honest and uncomfortable. The per-seat tools (Cursor, Linear, Notion, Loom) sit safely below the line because their AI workloads are bounded. The hybrid-usage tools (Bolt, Replit, v0) sit at or on the line because they're explicitly designed as pass-throughs with a thin platform margin. Lovable sits in between because its credit model gives it room to throttle.

The takeaway: if your product is artifact-heavy, accept that you will run thin margin until inference prices fall. If it's seat-heavy, you have room. If you're sitting on the line and you're not VC-funded, you have a quarter to fix it before the cash runs out.

Decision tree: which model to pick

Walk this in order and stop at the first match.

1. Does your AI produce a discrete artifact per run (a build, a video, an image, a brief)? Pick usage-based or hybrid usage. Bolt and v0 shape.
2. Does your AI assist a human user who keeps coming back (notes, code review, search)? Pick per-seat. Linear, Granola, Loom AI shape.
3. Does your AI replace a job a human would otherwise do (close support tickets, write briefs, run a campaign)? Pick outcome-based. Intercom Fin, SoGood shape.
4. Is your AI cost per active user higher than 20 percent of any of the above prices? Layer hybrid: keep the flat fee, meter the AI on top.

The wrong question is "what do my competitors charge." The right question is "what is my AI COGS per active user, and what is the minimum price that keeps me above 50 percent gross margin."

Margin math: calc your AI COGS per user

This is the work most founders skip and then regret. The formula is straightforward; the inputs need real data, not vibes.

Step 1. Measure tokens per active user per month. Pull from your provider dashboard (OpenAI, Anthropic, Together). Separate input and output. Get the median and the 90th percentile per user, not just the average. The 90th percentile is the user who will break your margin.

Step 2. Multiply by your blended model price. Frontier models in 2026 run roughly $3 to $15 per 1M input tokens and $15 to $75 per 1M output tokens, depending on which model and which provider. Use your actual blend. Cached reads are 80 to 90 percent cheaper; bake that in.

Step 3. Add non-LLM infrastructure cost. GPU time for self-hosted models, vector DB cost, storage, bandwidth. For most LLM-API products this is 10 to 30 percent of the LLM bill. For RAG-heavy products it can match the LLM bill.

Step 4. Divide by active users. This is your AI COGS per active user per month. Compare to your headline price.

Step 5. Sanity-check at the 90th percentile. Take the same calculation for your top-decile user. If that user costs more than 1.5x what they pay, you have a power-user problem and need a credit ceiling or overage. The supporting math discipline is the same we'd run inside a business plan; we cover the tool side of that in AI business plan generators honest review and the books side in AI bookkeeping software for startups 2026.

What goes wrong

Anchoring to ChatGPT Plus. $20 per month is OpenAI's subsidized anchor, not a market price. Copying it without OpenAI's enterprise revenue underneath is how bootstrapped AI SaaS dies in 9 months.

Unlimited tiers in artifact products. Unlimited builds, unlimited videos, unlimited generations is a marketing line that becomes a margin death sentence the moment one user discovers the loop. Bolt and Lovable both retreated from unlimited language in 2025.

Pricing before instrumenting. You cannot price what you cannot measure. Ship usage telemetry per user before you set a price ceiling. The median is not enough; you need the distribution.

Outcome pricing for fuzzy outcomes. Outcome-based works when the outcome is binary and countable (ticket closed, brief shipped). It fails when the outcome is subjective ("a good summary") because attribution disputes will eat your support team.

What to do this week

1. Pull your token-per-user distribution for the last 30 days from your provider dashboard. Get median, 75th, 90th percentile.
2. Compute AI COGS per active user. If it's above 30 percent of your current price, you have a quarter to fix it.
3. Walk the decision tree above. Pick one of the four models, not a hybrid of all of them.
4. Set a hard cap or overage path before you change the headline price. Cap first, then re-price.
5. Publish the cap and the unit clearly. Buyers forgive limits they understood at signup; they hate limits that surface mid-month.

If you want a bundle that does this thinking for you on the way to shipping a business, SoGood is outcome-priced: $29 per month Pro for a website plus a business case, $99 per month Expert adds ads, newsletters, and the rest of the 8-department stack. We're not a pricing tool. We're an example of an outcome-priced AI bundle that ships work, which is one of the four models above. The pillar context is what is an AI cofounder.