The most expensive mistake we made in 2024 was pricing PrepareMesCours like a vintage SaaS: a flat 9 EUR per month with unlimited generations. A teacher signed up, generated 800 lesson plans in 11 days for an entire school year of prep, and we lost money on the relationship for the next 14 months. We do not blame the teacher. The pricing was wrong. Pricing AI features SaaS founders ship in 2026 needs to handle real LLM costs without making the customer feel like a metered taxi. This is the model we converged to across four products and the reasoning behind every number.
What "scaring SMBs away" actually means
SMB customers (teachers, retirement advisors, single-cabinet professionals) have three pricing fears that are not the same as enterprise:
- Surprise bills. A 9 EUR plan that becomes a 47 EUR bill because they "used too many tokens" feels like a betrayal.
- Per-feature paywalls. "Oh, that is only on the Premium tier" repeated three times in a session is the fastest way to lose a customer who was 10 minutes from paying.
- Pricing pages they cannot read. If your pricing page has 14 features in a comparison grid and the customer needs 90 seconds to figure out what they get, you have already lost them.
Our pricing has to be the opposite of these three things. Predictable, generous within reason, and readable in 6 seconds.
The unit economics, with real numbers
Let us put real costs on the table because most pricing posts skip this and it is the entire decision.
For PrepareMesCours, a single lesson-plan generation costs us roughly:
- LLM call cost: $0.04 to $0.09 per generation depending on the prompt complexity and the output length. We use OpenRouter routing to a model that delivers the right quality at the right price.
- Image generation (when used): $0.01 to $0.03 per image, used in maybe 30% of generations.
- Storage (the generated lesson goes to MinIO on our own hardware): rounding to zero.
- Database write and read: rounding to zero.
So a lesson plan costs us roughly $0.07 average, all-in. Call it 7 cents.
A teacher on a typical month generates around 25 to 60 lesson plans. That is 25 to 60 times $0.07, or $1.75 to $4.20 in cost-of-goods per month. At our current 9 to 14 EUR pricing, the unit margin is healthy.
The problem is the outlier teacher who generates 300 plans in a month. That is $21 in costs at the upper end. If that teacher is on the 9 EUR plan, we are losing money. Hard.
The first job of AI pricing is not to capture upside. It is to prevent the long tail from eating your margin.
The pricing model that worked
After three iterations, here is what we landed on across the lesson-planning products:
A monthly fee with a generous-but-finite cap
Everyone says "soft caps". We tried that. Soft caps are unenforced caps. Customers do not believe them. We use explicit hard caps with a clear number on the pricing page: "60 generations per month included".
Most teachers use 25 to 40 per month. The cap is generous for normal usage. The 5% who hit it can buy a top-up pack (10 extra generations for 3 EUR) or upgrade to a higher tier with a higher cap.
The caps are visible in-app at all times. A small progress bar in the dashboard shows "23 / 60 used". Nobody is surprised.
Three tiers, not five
Free trial (5 generations, no card). Standard (60 generations per month, 11 EUR). Pro (300 generations per month plus exports and team features, 24 EUR). That is it. No "Enterprise call us" tier on the pricing page (we have one, it is just not on the page).
Three tiers fit on a page. Five tiers do not. SMB customers do not pick the middle tier "for a reason". They pick the middle tier because they cannot read five columns at a glance.
One headline number per tier
Our tiers each have one big number that the customer instantly understands: 60 / 300 / unlimited (with fair use language). The other features (export PDF, image generation, custom branding) are in a smaller list. The customer's brain anchors on the headline.
We tested this against the more conventional "feature grid" layout in late 2025 and the headline-number page converted around 1.4 times better. We are not above using a feature grid. We are above making it the primary visual.
The Carriva pricing exception
Carriva sells to retirement advisors, not teachers, and the pricing model is different. Advisors pay around 79 EUR per month per seat with included audits, and they expect to pay more than a teacher would. They are running a business that makes money from advice. The price needs to be visible (we hid it once and lost leads) but it does not need to be cheap.
The lever for Carriva is not price. It is onboarding speed and the free B2C detector that brings end-clients into the funnel. We discussed the broader strategy in our piece on why we built Carriva and the vertical AI SaaS thesis we apply to it.
The lesson is that pricing AI features SaaS founders set should match the buyer's economic context, not the model's cost. A teacher and a retirement advisor have different willingness-to-pay even though both are SMB.
What we tried that did not work
For honesty:
Pure usage-based pricing
In our first version of DraftMyLesson we tried "pay per generation, $0.40 each". It did not work. Teachers would do mental math at every click. They wanted predictability and they wanted to feel safe. Conversion fell off a cliff. We pulled the model in 6 weeks.
The lesson: usage-based pricing works for developer tools (where the buyer is sophisticated) and for enterprise (where the buyer has a budget code). For SMBs it adds friction at the worst possible moment, the moment they would otherwise generate the artifact you want them to generate.
Token-counting visible in the UI
Briefly we showed customers how many tokens their generation used. We thought it would build trust. It built anxiety. The same teachers asked us "is 1,800 tokens a lot?" and the answer ("no") did not help. We removed it.
A tiered "AI quality" pricing
We considered offering a cheaper tier that uses a smaller model and a premium tier that uses the best model. We did not ship it. The reason is that customers cannot evaluate "AI quality" in the abstract. They can evaluate "did this lesson plan work in my classroom?" The pricing should not put a quality lever in their hands when they cannot judge what the lever does.
The hidden costs we built into pricing assumptions
A pricing page is a forecast. It assumes things about your costs that may not hold. We learned to budget for:
- Model price changes. Provider prices change quarterly. We assumed a 20% margin buffer to absorb a price hike without changing customer pricing immediately.
- Edge cases. That one teacher who runs the system through a script and generates 1,200 outputs in 2 days. We have a fair-use clause and we have invoked it twice in 18 months.
- Refund risk. Some customers churn in week 2 and ask for a refund. We give it. The cost is real but small. We budgeted around 4% of revenue for refunds and we are running closer to 2.5% in practice.
- The free tier as marketing spend. A 5-generation trial costs us roughly $0.35 per signup. With trial-to-paid conversion at 9%, that is around $4 in customer acquisition cost from trials. Cheap by any standard. But it is a real line item.
The unspoken pricing principle: do not be greedy
The temptation in AI SaaS pricing is to capture every marginal dollar. We push back on that. The reason is that SMBs talk to each other. A teacher who got hit with a surprise upcharge tells five teachers in their staff room. A retirement advisor who felt squeezed tells their professional association.
Our principle: be transparent, generous on normal usage, and firm on outliers. Most customers feel respected. The few who try to extract maximum value at our expense are politely capped. The cap is the cost of running a sustainable studio.
We feed pricing decisions into our blog and customer-development cycle. When a prospect asks "is this expensive?", we know it is a positioning issue and not a price issue (we have done the math). When ten prospects ask "I cannot tell what I get for the price", we redesign the page. Customer feedback drives the page, not our gut.
For context on how we generate the marketing content that shows up on these pages, we wrote about the cost of AI-generated content for SaaS marketing including the real per-article numbers.
A small studio's pricing checklist
If you are designing or rebuilding pricing for an AI SaaS aimed at SMBs:
- Calculate the actual cost-of-goods per use. To three decimal places.
- Set the included usage cap at roughly 4 to 6 times the median customer's normal usage. Cover the 95th percentile. Cap the 99th.
- Show the cap in-app, not just on the pricing page.
- Use three tiers, not five.
- Use one headline number per tier.
- Build a 20% buffer for model price changes.
- Decide how generous you want to be on the trial. Most SMB SaaS over-restrict trials and lose conversion.
- Re-price annually. Not monthly. Annually. Customers can absorb a yearly review. They cannot absorb random monthly changes.
What is next
For us, the next pricing experiment is bundle pricing across the lesson-planning products: a teacher who teaches in two languages (some do, especially in international schools) gets a discounted dual-locale plan. We will measure conversion and retention impact for two months before deciding. If it does not move the needle, we will not keep it.
Pricing AI features SaaS founders ship in 2026 is more art than spreadsheet, but the spreadsheet is non-negotiable. Run the math, watch the long tail, and respect the SMB customer's relationship with risk. If you do those three things, the rest is just iteration. Customers who feel respected pay you for years. Customers who feel surprised churn in 30 days. Pick the long game.
