Peter Deng on Why AI Pricing, Evals and Voice Are Still Misunderstood

Peter Deng says subscriptions will evolve dramatically in AI, and he roots that claim in his first project at OpenAI: ChatGPT for Enterprise. He led consumer products, product marketing and design, wrote the first spec and vision document, and the team launched in seven weeks while he was quarantined with COVID. During pricing calls he felt something was wrong with per-seat pricing because the unit of work is not access to a tool, as in classic SaaS, but the work provided and done.

He argues that per-seat pricing can be set high enough for heavy GPU users, but it still misaligns value because some customers are undercharged and others overcharged. Usage-based pricing is not the answer either, because tokens are a cost driver rather than value; he points to startups using a platform fee plus a usage fee and working backward from the efficiency and impact they create. Buyers have a mental hangup about outcomes-based pricing, but he thinks they can accept a platform-plus-usage model once they see the outcomes after a year.

On limits, Peter says the product should let 80% to 90% of users never hit a constraint, and when they approach or exceed one, the experience should gently guide them to the next tier rather than make them feel they did something wrong. Seat-based pricing works for a broad product like ChatGPT because most users can believe they have all they can eat under the limit, which creates a natural upsell path. For enterprises, he expects a broad platform fee plus power-user usage and admin tier choices, and he strongly advises against pricing purely on usage or GPU time because that turns the product into a commodity race to the bottom.

Jacob Effron asks how product teams should work with research and post-training, and Peter describes the PM as a service organization that fills the white space between disciplines. At Facebook and Instagram the core was a triad of product manager, designer and engineer; at Uber it added an operations person because the product operated in the physical world. At OpenAI the configuration changed again to product, design, engineering and the post-training lead, because the model is the product.

Peter says the PM job shifts because the new spec is the eval: agent behaviors and boundaries cannot be captured in prose alone. The post-training lead, product lead, design lead and engineering lead must operate as one unit, with the PM picking up any piece that is not being picked up. Jacob notes that evals are the big new piece in this wave, and Peter agrees emphatically.

Peter sees RL and tool use getting good enough to open a new abstraction layer for startups, especially because much valuable data is not on the open web and models are now malleable enough to absorb it. Founding teams need to have played with this new element and gotten it under their fingernails; the best builders understand the raw materials even if they are not researchers. Model companies have early access and can tune behavior with post-training, but startups can dedicate 100% of their time to a specific use case and move fast.

Focused startups can build tailored products for broken workflows that ChatGPT cannot handle, and bespoke post-training against the evals they care about differentiates them from horizontal players. Peter says both sides have advantages: labs can make many boats rise, while startups can go deep on a market insight and move incredibly fast. He has not seen pitches that sit right down the fairway for OpenAI or Anthropic; trillion-dollar buckets may attract the labs, but billion-dollar markets can be too small for them and still large for startups.

Peter frames any role by where people spend their mental horsepower, treating attention as a limited compute budget. Specialization will persist because someone must own the craft of code and architecture, advocate for design simplicity and insist on engineering robustness, even if code drips from Cursor. Every working human is increasing the level of abstraction they operate at, but ownership and advocacy do not disappear.

He calls ChatGPT the PM's best friend: a late-night thought partner that is never tired and can help crack the last word of a spec or vision doc. Every role gets more leverage, and PMs get a round-the-clock partner. Jacob asks whether that means more or fewer PMs; Peter says roles change constantly and predicts hybrid solo builders who can design, code, operate and ship a feature without the old PM/engineer split.

Peter cites an OpenAI engineer he admired who had gone to design school, and he says AI lets people supplement and complement their skills to fulfill a vision. The PM title may persist but become more technical, or it may not. He notes that PM archetypes have already shifted within two decades, from more marketing-oriented roles at Apple and Microsoft to the technical, design-oriented PM at Facebook, so another shift is plausible.

Jacob asks how voice fits into consumer AI, and Peter points to the original voice-mode launch sticker on his water bottle. He studied human-computer interaction and says each modality has its own pros and cons, but voice unlocks something because people start and stop thoughts while speaking; the brain forms the thought and the words at the same time, with no lag. Voice models are also getting better at tonality and warmth, and he calls Sesame's demo magical and finely crafted.

Jacob wonders which voice use cases will persist, especially if voice agents end up calling other voice agents, which may be less effective than other channels. Peter still sees voice as an incredibly engaging medium that avoids the blank-box problem of formulating a thought before typing it. He uses Granola on drives to talk through an investment thesis with false starts, then gets a structured agent-processed summary showing which insights emerged and when. Voice has stuck with humanity for a long time, and he expects it to remain important.

On education, Peter says many startups are forming and the models can be told not to give the answer but to help someone think through it. He has used ChatGPT to remember long division and to help his kids with homework when his brain was fried late at night. Jacob mentions language learning and Sal Khan's point that engagement is the hard problem; Peter says AI can follow kids' passions, such as horseback-riding camp or framing math around soccer, to unlock curiosity and a lifelong love of learning.

Peter's favorite discovery at OpenAI was what happens when the model talks first, shifting it from an assistant on call to something proactive. He points to Utori as a company already alerting users proactively, and he sees much more room in that arc. Jacob notes that the blank box works for the roughly 10% who know what to do but can confuse many others, so a proactive shift is especially interesting.

Jacob asks how diffusion compares with expectations, and Peter says OpenAI hired him for both consumer and enterprise because ChatGPT is a work product. Enterprises needed security, integrations, audit logs and privacy, so adoption had to wait for those basics and mental models to be built. On the consumer side, fear was the early consensus; a product marketing hire called users in the Midwest, and over the last 18 months playing with the technology has dissipated some fear and made the usefulness evident.

On the research talent wars, Peter trusts the market because research talent is critical for each big lab to retain and recruit. He thinks the winners are everyone else, including investors and startup founders, because competition toward general intelligence and superintelligence is good for humanity. He also flags a second-order effect: non-researchers who contribute heavily to products and companies see salary and RSU gaps widening, and HR departments will eventually have to address it.

Jacob links talent pricing back to agent pricing: if agents replace labor but all do similar work, they may compete each other down, and if they are meaningfully better for a role, they can keep pricing power. Peter says companies must model two or three degrees ahead and find the next level of abstraction. He uses Facebook as a springboard example, moving from a college site to global communication to advertiser reach and now hardware plus superintelligence, crediting Mark Zuckerberg with seeing around the corner.

Peter describes Facebook as a series of springboards: a college site, then high school and global communication, then advertiser reach, then hardware and superintelligence. He respects Mark Zuckerberg's ambition and his ability to bet on the next platform before it is too late. On Meta's AI strategy, he believes the 'personal super intelligence' framing and expects many startups to race for pieces of that vision.

His hot take is that the word 'agent' is overhyped, because today's systems can think, talk and see but lack the tools to make a reservation or rearrange a schedule. Everyone is building toward that holy grail, but nobody has fully accomplished it. Jacob asks whether there will be one consumer agent or many, and Peter says many, because humans demand variety and products break down by jobs to be done, workflows, unique data and orientation; big companies will leverage distribution points such as Chrome or feeds.

On defensibility, Peter starts with unique data that is not on the web, such as healthcare scheduling transcripts and tonality, the timing of pauses in sales cold calls, or internal research and chain of thought at firms like McKinsey or Bain. He reminds founders that Facebook's unique data was mapping who people were really friends with, which had lived only in people's heads and became a huge company once digitized. The second layer is a data flywheel from the right workflows, and the third is speed to blanket the market before competitors catch up.

Peter sees deployed users as mini-labelers: when an agent mishandles a scheduling task and the user corrects it, that correction is unique data available only to the entrenched tool. Some of that learning updates the model, and some updates the harness and workflow around it. Jacob notes that models need surprisingly few examples, and Peter agrees that fine-tuning on around 100 data points can start a flywheel; the deeper value of deployment is discovering edge cases and scaffolding. He adds that behavior shifts over time, as consumers moved from landscape to vertical video, so the flywheel also helps detect changing preferences.

Jacob asks about GPT-5's mixed reception, and Peter credits the whole OpenAI team while taking zero credit himself. He says the backlash was not a surprise, because Facebook's homepage and News Feed revamps triggered the same 'you moved my cheese' reaction. Consequential products with deep followings go through change management, and he saw the GPT-5 reaction as the same pattern. The direction toward no model picker is right, but after an earlier cleanup the picker grew back because people want control.

Peter says product design has to balance magic and friction removal against how much control users need and who needs it, citing News Feed and Twitter as examples of different desired control levels. Jacob asks about ChatGPT as a consumer homepage versus Uber or DoorDash integrations, and Peter frames it as an incentives problem: some company will partner more deeply, which will force others to follow. Everyone optimizes their own business, but the world is moving toward simplicity and instantaneousness, so Uber versus Lyft will come down to ETAs, service quality and how well an agent can access them.

Peter predicts interfaces will largely go away, with devices in the ear that are proactive rather than reactive, and he expects different players to own different pieces before things collapse into a voice model or device interface. That is why Apple and Google have an advantage and why other companies are pursuing hardware. Jacob asks whether this is net bad for Uber; Peter notes that the iPhone and GPS gave Uber its chance, and super apps already exist in parts of Asia, so a voice super app is plausible and Uber will have to adapt and differentiate to avoid commoditization.

Jacob cites Ben Thompson's writing on OpenAI's consumer-versus-enterprise tension, and Peter says he is glad not to have to solve it, though he trusts the applications leadership to figure it out. At OpenAI his title was hard because the scope was not just consumer or enterprise but also API and platform, and enterprise might include API too. When he left, enterprise and consumer ChatGPT were split into two groups; he later heard they were back together, which makes sense because it is the same product. He sees reorgs as a way to express strategy, so once the next 9 to 12 months of strategy are clear, the org will match.

Jacob contrasts OpenAI's breadth with Anthropic's sharper focus, and Peter says OpenAI has not missed over the last few years. He recalls SearchGPT as a separate entry point that rallied the organization around web search by default and was later folded into the main product; he frames that not as failure but as finding a better way. Pre-product-market-fit startups should learn from experiments and pivot, and he sees adaptability and a growth mindset in every generational company.

Peter says OpenAI dreamed bigger than any other company he has worked at, and its all-hands left his brain stretched by the scale of AGI, research agents that self-perpetuate and keep making advancements. He is an optimistic person, though as an investor he sometimes has to tone optimism down; OpenAI stretched his optimism and inspired him. In the quickfire round he calls agents overhyped today because they do not yet do what agents should do, and evals underhyped because they are the unsexy elbow grease that unlocks model behavior. If he were starting an AI app, he would go into a boring business with wasted human effort, especially AI voice and call centers.

On parenting, Peter says he wants to foster curiosity and the boldness to believe that whatever his kids dream up, they can realize. He expects they will have tools beyond current imagination and is excited about the creativity they will bring. He still believes fundamentals matter: math teaches structured thinking, and coding teaches how to take a problem and subdivide it into component parts, which is why Cursor is a blessing rather than a threat. He closes by asking for warm intros to founders with clarity, grit, ambition and elbow grease.