How to control AI costs and pricing

1 The bottom line

Pick the model by what the user's value is worth, spend freely while you learn, and stop paying for users who have not paid. Costs fall fast, but trial length, free tiers and daily features can still lose money.

2 Key takeaways

• Build for quality first, because model prices keep falling. Cheaper models arrive fast enough to fix unit economics later.
  • After a new model launched at 50% cheaper, another arrived two or three weeks later at 50% cheaper again Adam Turaev, Praktika55:13.
  • Praktika earns more than $1M revenue per employee, with about 40 people and $50M ARR Adam Turaev, Praktika55:13.
• Match the model tier to the upside of the job. Top intelligence is very expensive, so it only pays where the upside has no ceiling.
  • One extra IQ of intelligence costs 100X more than the prior model Anish Acharya, a16z22:00.
• Test whether a feature is affordable for every user, every day. Small features add up fast at scale.
  • RevenueCat estimated that one LLM call per app per day would cost tens of thousands of dollars a month Andrew Maguire, Volo Ventures29:20.
• Count trial length as a cost, not only a conversion lever. Every extra trial day is paid inference for people who may never convert.
  • In the US, 7-day versus 3-day trials showed about a 20 cent ARPU difference Morrie Schonfeld, Pingo53:00.

You need to make sure that the value that you're creating is actually higher than the cost that you incur.

Kenneth Schlenker, Opal

2.1 Check the economics before you build

Decide whether the value beats the cost before a feature reaches every user.

• Ask if you can afford the feature for every user every day. Source: Andrew Maguire, Volo Ventures29:20
  • Run this test before you commit, because even small features add up at scale.
  • Opus pricing has not come down in six months Andrew Maguire, Volo Ventures33:30.
• Compare the value you create with the cost before you set the free tier. Source: Kenneth Schlenker, Opal20:00
  • Server-side AI models carry a marginal cost per user, which changes the freemium equation.
  • After the excitement of building, ask whether you can charge enough.
  • Test with real users before you decide what is free.
  • An AI coaching feature got expensive, and the model went from Opus 4.6 to Sonnet to Gemini Flash Kenneth Schlenker, Opal20:00.
• Check whether token use tracks value. Source: Connor Zwick, Speak31:42
  • Tools where per-token usage is highly correlated with value face the most cost pressure.
  • A premium subscription with modest token use per learner has more breathing room.
  • That margin lets Speak stay on the bleeding edge of new models Connor Zwick, Speak31:42.

2.2 Spend freely while you learn

Early on, finding out what works matters more than saving tokens.

• Let engineers use as many tokens as they want at the start. Source: Boris Cherny, Claude Code26:48
  • Throw a lot of tokens at an idea and see if it works, so people try ideas that would seem too crazy otherwise.
  • Optimize once the idea works and scales, for example by moving to Haiku or Sonnet instead of Opus.
  • Some Anthropic engineers now spend hundreds of thousands of dollars a month in tokens, and some companies offer unlimited tokens as a perk Boris Cherny, Claude Code26:48.
• Treat early token spend as part of acquisition cost. Source: Brett Bauman and Zack Hargett, Coconote20:50
  • Coconote paid little attention to LLM costs early and treated them as a good problem to have.
  • It used startup credits, then made cost-efficiency decisions as it grew.
  • Audio transcription cost more than the LLM.
  • Each founder invested about $10K up front, and the company ran around 50% EBITDA margins nearly every month until the acquisition.
• Tell teams not to hold back over cost when a feature shows ROI. Source: Connor Zwick, Speak32:30
  • Assume costs will keep going down, and work to dispel cost-based self-constraint.
  • Speak has far less margin pressure than agent coding companies, though multimodal speech is still expensive.

2.3 Match the model to the job

Pay for top models where quality earns money, and use cheaper ones where it does not.

• Use frontier models for unbounded work and cheaper models for bounded work. Source: Anish Acharya, a16z22:00
  • Top intelligence makes sense in drug discovery, sales, support, research and engineering.
  • For bounded jobs like closing the books, use open-weight models tuned with reinforcement learning.
  • Expect companies to run both architectures, since frontier tokens are irrationally priced on the Pareto frontier.
• Choose the model for your customer. Source: Phil Carter, Elemental Growth49:50
  • Product experience depends on output quality, speed and cost.
  • A cheaper, faster model can be good enough for price-sensitive users. If users pay premium prices and peak performance matters, use the best model.
  • Plan for 6 to 12 months out, because API prices are falling.
  • Gamma was profitable 6 months after its March 2023 viral launch, helped by savvy model choices.
• Test which model each subscription tier gets. Source: Jeff Morris, Tinder3:22
  • Decide whether a tier gets your best, more expensive model or cheaper ones that can still convert.
  • A startup with a $50 million Series A can offer a better experience with an identical product, so cost shapes your options. Non-AI products can use the classic playbook.

2.4 Cut waste in the free tier and the trial

Do not pay inference costs for users who may never convert.

• Run the expensive API only after the user pays. Source: Nicole, Glam Up36:40
  • Glam Up showed a fake, blurred face scan result first. The real API call ran after payment.
  • This avoided API cost for non-converting users.
• Route chat queries to cheaper models so a free tier is affordable. Source: Ajay Mehta, Tolan46:27
  • Send many queries to cheaper models and infrastructure at the same quality, and monitor quality closely.
  • This lets you give more usage to a larger group of people.
  • Tolan only felt confident offering a moderately generous free tier earlier this year, once open source models cost a fraction of frontier models. Cost work came before conversion work.
• Weigh trial length against usage cost. Source: Ajay Mehta, Tolan6:51
  • When choosing between a three or seven day trial, remember the extra four days add meaningful cost for trialists who do not convert.
  • Pingo compares usage in the first 3 days with days 4 to 7, and factors in CAC, to see if the 7-day trial is worth it Morrie Schonfeld, Pingo53:00.
  • Pingo's 3-day cancel rates ran 5 to 7% higher in the US, but users use the app heavily in the first 3 days Morrie Schonfeld, Pingo53:00.

2.5 Pay for the cost with revenue or funding

Inference costs need a source of money, whether a paywall or funding that bridges the gap.

• Add a paywall before token costs and investors force one. Source: Koby Conrad, Sunflower43:59
  • Sunflower copied Duolingo's free-first approach, expecting to find revenue later.
  • Investors wanted to see money, so the team added a monthly paywall.
  • AI token costs were scaling to hundreds of thousands of dollars.
• Use venture money only when growth is viral and inference cost is the main gap. Source: Andrew Maguire, Volo Ventures33:30
  • The fit is a product with network effects, fast growth and addicted users, but high inference costs today.
  • The bet is that costs come down as long as growth continues.
  • Timing matters, since a cheaper model in two years versus five years gives very different outcomes.

2.6 Where they disagree

The right move depends on how fast costs really fall, and on how much margin your product has.

• Best model or cheaper model. Boris Cherny says to run the most capable model at maximum effort.
  • A less intelligent model like Sonnet often takes more tokens because of more correction and hand-holding, so the best model can be cheaper and faster in the end Boris Cherny, Claude Code1:09:00.
  • Tolan routes many queries to cheaper models at the same quality Ajay Mehta, Tolan46:27, and Gamma started with a cheaper, faster one Phil Carter, Elemental Growth49:50.
  • Andrew Maguire warns that cheaper models often fail on trivial tasks once you have used the best model, so routing is not a free fix Andrew Maguire, Volo Ventures29:20.
• Optimize cost early or late. Coconote and Speak say not to constrain yourself, while Andrew Maguire and Kenneth Schlenker say to test the economics first.
  • Speak can build freely because its premium subscription leaves room in the margin Connor Zwick, Speak32:30.
  • David Barnard pushed back that premature cost optimization can cost you the best experience Phil Carter, Elemental Growth49:50.
• How much to lean on falling prices. Adam Turaev expects model prices to fall and balance the economics later Adam Turaev, Praktika55:13.
  • Andrew Maguire agrees costs will fall, but says to be honest about how fast Andrew Maguire, Volo Ventures29:20.