How OpenAI Builds for 800 Million Weekly Users: Model Specialization and Fine-Tuning
Summary
OpenAI’s distribution strategy is deliberately two-sided: ChatGPT reaches roughly 800 million weekly users, while the API may touch an even broader population through third-party products. Sherwin Wu says this follows the founders’ mandate to distribute AI’s benefits across as many surfaces as possible, despite customers’ recurring fear that ChatGPT might enter their market. At its current growth rate, the first-party product is such a force that internal competitive anxiety is limited: “10% of the globe uses it every week.”
Models are proving unusually resistant to commoditization because both users and developers form model-specific dependencies. Users form preferences among GPT-5, GPT-4o, and o3, while developers tune prompts, tools, and application harnesses until they are effectively “building a product around the model.” That combination produces surprisingly high API retention and weakens the classic platform risk that an application abstracts away its underlying supplier.
OpenAI has moved from expecting “one model that rules them all” to increasingly expecting room for a portfolio of specialized models. Cursor illustrates the emerging stack: GPT-5 in Max mode for planning, Composer for fast first passes, and other models for deeper bugs or tab completion. Sherwin expects “a proliferation of other types of model,” though he cautions that today’s view may change again; specialization is not presented as a retreat from AGI, but as evidence that the path is less predictable than assumed.
Reinforcement fine-tuning turns proprietary enterprise data from a retrieval asset into potentially differentiated model capability. Earlier supervised fine-tuning mostly delivered “instruction following plus” and tone changes; RFT can potentially produce a state-of-the-art model for a customer’s narrow use case. OpenAI is piloting discounted inference and potentially free training for customers willing to share valuable datasets—an explicit exchange around the “giant treasure troves of data” sitting inside companies.
The application bottleneck has shifted from clever prompts to context engineering and controlled orchestration. Better reasoning models still need the right tools, data, retrieval timing, and procedural boundaries; Martin Casado’s reductio is that classic RAG uses cosine similarity to choose what gets fed “into a super intelligence.” For enterprises, intelligence does not eliminate workflow structure: many jobs require deterministic SOPs, approved response catalogs, and enforceable limits on what an agent may do.
OpenAI treats agents as intelligence expressed through different interfaces, not as a separate product category. ChatGPT, Codex, Sora, and the API are “different interfaces and different ways of deploying” the same core intelligence; Codex became separate because a CLI suited coding better than ChatGPT. This creates room for startups to embed or build around Codex—the economy Casado jokingly calls “token laundering,” turning natural-language input into metered intelligence.
Usage-based API pricing has held because it maps cleanly to consumption, supports cost-plus margin discipline, and may already approximate outcome pricing. OpenAI sees usage pricing as a possible “one-way ratchet”: once customers pay closer to realized utility, deployment-style pricing becomes hard to restore. Outcome billing remains appealing but difficult across domains such as crop yields; Martin argues that as test-time compute rises with task difficulty and value, token usage may become a workable proxy for outcomes.
Open weights appear additive rather than cannibalistic, while efficient inference remains a meaningful operational moat. Sherwin says OpenAI has “not seen cannibalization at all” from gpt-oss because its customers and use cases differ from those of the major hosted models. Even hypothetically releasing GPT-5’s weights would not reproduce OpenAI’s serving performance: its largest two or three models receive concentrated resources, and the inference and training teams optimize together. Multimodal breadth likewise depends on specialized stacks—shared API infrastructure above separate text, image, and video inference systems.
Deep dive
1. OpenAI distributes intelligence through both product and platform
Sherwin’s operating lens spans Quora feed ranking, six years pricing Opendoor’s housing inventory, and roughly three years working on OpenAI’s developer platform. That remit is primarily the API, but also includes government deployments such as a model running locally on a classified supercomputer at Los Alamos National Laboratory.
The apparent product-platform conflict is framed at founder level. Sherwin says Sam and Greg have been consistent “since day one”: OpenAI wants ChatGPT as a first-party application and the API as a broad distribution channel, both serving the mission of distributing AI’s benefits across as many surfaces as possible.
ChatGPT’s scale is the clearest first-party proof: about 800 million weekly users, or “10% of the globe,” with continued growth. Sherwin calls the speed and operational scaling historic; Martin emphasizes that many listeners may not grasp how extraordinary that weekly reach is.
The API’s indirect reach is harder to compare and, at points, was “definitely larger than ChatGPT,” though Sherwin hedges that ChatGPT is growing too quickly for a stable comparison. Tension is felt mainly by API customers asking, “Are you going to build this thing that I’m working on?” ChatGPT itself is not particularly worried because growth remains so strong.
2. Model identity makes the API harder to disintermediate
Martin’s anti-disintermediation thesis: unlike conventional infrastructure, models are too “unruly” to hide cleanly behind a software abstraction. If users must encounter the model’s behavior directly, an application cannot make its supplier invisible as easily as software built on a fungible cloud service.
Sherwin says the industry originally expected models to be interchangeable infrastructure. Product reactions to GPT-5, GPT-4o, and o3 challenged that assumption: users recognized distinct behavior and formed preferences. Martin felt the change personally—“You’re not as nice to me”—while Sherwin says he loved GPT-5’s personality but used ChatGPT mostly for work or information.
The technical dependency runs deeper than sentiment. Startups iteratively optimize prompts, tools, and application harnesses around one model; as that work compounds, they are “building a product around the model,” and switching threatens the product’s model-specific behavior as well as user familiarity. API retention has therefore been “surprisingly high,” even where tooling makes substitution possible.
Martin’s own Cursor workflow is the specimen: GPT-5 is his default, Max mode with GPT-5 handles planning, Composer supplies fast first passes that preserve flow, and another model comes out for a difficult bug. The product’s value comes partly from exposing model choice rather than pretending all intelligence is equivalent.
3. The “one model” thesis has given way to specialization
Sherwin remembers that only two or three years earlier, even OpenAI expected “one model that rules them all.” Under that premise, fine-tuning barely made sense: a universal model would eventually subsume every customized capability, matching the simplest picture of how AGI would arrive.
That view has “definitely completely changed.” OpenAI now operates GPT-4, GPT-4o, GPT-5, Codex-oriented models, and other variants because distinct workloads reward distinct behavior. Sherwin increasingly expects room for specialized models and “likely” a proliferation of model types, while explicitly allowing that the industry’s view may change again.
Martin argues this could produce a healthier ecosystem rather than winner-take-all consolidation. Sherwin agrees that a growing AI market generally benefits OpenAI: its evals product accepts other models, customization supports differentiated variants, and gpt-oss broadens participation. He does not treat specialization as bad for AGI—only as another unexpected turn in pursuing it.
4. Reinforcement fine-tuning turns private data into capability
Fine-tuning demand begins with companies’ “giant treasure troves of data” and their desire to participate in the AI wave. Retrieval through a vector database or RAG is the simplest use; technically stronger teams want the data to modify model behavior, not merely appear in a context window.
OpenAI’s 2022–23 supervised fine-tuning offering was too limited for much of that ambition. Sherwin characterizes SFT as “instruction following plus”: useful for tone and more reliable obedience, but generally only an incremental change rather than a major increase in domain capability.
Reinforcement fine-tuning changes the proposition. It is more finicky and requires greater investment, but it can exploit domain data far more deeply—potentially reaching state-of-the-art performance on a specific use case. The sales pitch becomes not “make the model speak differently,” but “create the world’s best model using your dataset.”
OpenAI imagined LoRA-based infrastructure scaling to “millions and millions” of fine-tuned models, including the kind of online-learning setup that would support continual updates. In practice, most customers use offline datasets accumulated historically or produced by experts. OpenAI is also piloting discounted inference and potentially free training when customers voluntarily share valuable RFT data; the choice remains with the customer.
5. Prompt engineering has become context engineering
Another discarded 2022 assumption was that scaling would make prompt engineering disappear: instruction following would become so strong that users could simply state an intention and “mind meld with the model.” Models have improved enormously since GPT-3.5, but Martin’s summary is blunt: “That’s not a thing.” Sherwin agrees that the original belief was clearly wrong.
The work has migrated toward context engineering—choosing tools, deciding what data becomes available, and retrieving the right material at the right moment. The question is less how to word one magical instruction and more how to construct the environment in which a reasoning model can operate reliably.
Martin’s critique of classic RAG is deliberately absurd: cosine similarity in an embedding space selects material to feed “into a super intelligence.” Sherwin notes that RAG arose before reasoning models, when systems had one weak shot; o3 changed the pattern by diligently making tool calls, retrieving information, and writing code to execute as part of its own reasoning.
6. Agents are behavior expressed through interfaces
Sherwin’s utilitarian definition of an agent is “an AI that will take actions on your behalf” and can work across long time horizons. OpenAI does not organize around agents as one singular modality; agentic behavior is a way increasingly capable intelligence manifests across each product line.
Sora, Codex, the API, and ChatGPT are therefore “different interfaces and different ways of deploying” intelligence. Each product area asks what agentic capability should look like on its own surface, rather than handing the concept to one centralized agents team.
Codex demonstrates the interface decision. OpenAI explored coding behavior inside ChatGPT, but concluded that a CLI better matched the work and turned Codex into its own product. On the platform side, the developer-first manifestation includes a Codex SDK; Sherwin says a startup built around it would be “great.”
Casado’s memorable economic framing is that much of the emerging economy is “token laundering”: convert English or another natural language into metered intelligence, then place that intelligence inside a larger program. The interface, cost structure, and defensibility differ from traditional software even when the model is only one component.
7. Usage pricing remains the cleanest proxy for utility
Sherwin’s honest explanation for API usage pricing is that it was simply “closest to how it’s actually being used.” Unlike ChatGPT’s subscription model, the API must meter many customers with large quotas; OpenAI now has an entire team maintaining the billing infrastructure required to do that correctly at scale.
OpenAI prices from a cost-plus perspective, explicitly protecting responsible margins rather than applying Opendoor-style asset-pricing sophistication. The operational challenge is substantial, but Sherwin thinks usage pricing has “surprisingly held strong” and may remain the model for a long time.
Relaying the Rockset founder’s formulation, Sherwin calls usage pricing a “one-way ratchet”: once buyers experience billing tied more closely to utility, returning to per-deployment pricing becomes unlikely. The cost is that the vendor must operate the metering, quota, and billing infrastructure reliably.
Martin’s pushback on outcome pricing is that it forces the vendor to value non-computing outcomes—porting code may be tractable, but “increasing crop yield” demands vertical expertise. Martin then observes that test-time compute may make usage highly correlated with value: prolonged reasoning consumes more precisely when the task is harder and potentially more valuable. Sherwin agrees that this may amount to “just adding more value.”
8. Open weights expand the market without obvious cannibalization
Sherwin says OpenAI always wanted to release an open model; the practical constraint was sequencing scarce compute against frontier work. He and Sam had discussed it since Sherwin joined, making gpt-oss a delayed expression of that intent rather than a sudden reversal from an anti-open-source position.
OpenAI’s strategic case is ecosystem expansion: more use cases and industries create a rising tide for its product and platform businesses. Sherwin reports “not seen cannibalization at all”; open-weight users, hosted-API customers, and their workloads have so far been meaningfully different.
Martin sharpens the terminology: these releases are open weights, not enough to recreate the entire training process. He argues that competitive enablement may be limited because, if API outputs can be distilled comparably, providing weights does not fundamentally change that dynamic. Meanwhile, Sherwin says reliable high-speed inference remains “a hard, hard problem.”
Sherwin says major labs concentrate impact, revenue, and optimization on only two or three models. Even if OpenAI hypothetically released GPT-5’s weights, matching its production inference would be extremely difficult; OpenAI’s inference team is unusually strong, and its feedback loop with training enables joint optimization that weights alone would not capture.
9. Multimodal breadth works because the stacks stay separate
Deeper text-model verticalization is beginning with GPT-5 Codex, particularly its post-training and ability to use the Codex harness. How far to push product-specific training remains “more of an open question”; today’s work is not yet the deepest possible fusion of model and application.
Image models permit faster specialization because they tend to be smaller and easier to iterate on. Text models retain a “really big fat pre-training step,” while full mid-training or post-training remains computationally heavy, limiting the proliferation seen on the image-model side.
Sherwin agrees that building leading language and pixel systems inside one company is an “anti-pattern” and difficult to execute. OpenAI manages it through a highly concentrated world-simulation team with a separate roadmap, productization process, and inference stack; Sora emerged from that organizational independence.
Above inference, the API infrastructure can be shared. Beneath it, text, image, and video models are served differently and optimized by their specialist teams. DALL-E 2’s generated white duck originally drew Sherwin to OpenAI; Sora 2 later became a “huge hit” in the API after Dev Day, surprising him with the reception.
10. Deterministic agents fit how enterprises actually operate
OpenAI’s Agent Builder prompted criticism because its connected nodes looked deterministic, low-code, and insufficiently “AGI forward.” Sherwin’s practical answer is that models still cannot execute every requested four-step process perfectly, while customers want useful automation now. Its launch around October 6 nevertheless produced one of OpenAI’s most-viewed demo videos and strong user enthusiasm.
Customer research exposed two categories of work. Coding and data analysis can be undirected: give an agent a high-level objective and let it explore. Much more enterprise work is procedural and SOP-oriented; Sherwin uses customer support as an example, while Martin says the pattern extends to marketing and sales. Managers often want policies followed exactly and do not want creative deviation.
Martin supplies the strongest regulated-industry example: some organizations cannot expose any generated language to users. They instead provide an approved conversation tree or response catalog and let the model select among valid outputs—using language reasoning while ensuring “nothing generated comes out.”
His NPC analogy makes the constraint concrete. A game may express inventory, barter, keys, and battle logic as Python-like functions because English descriptions are unreliable, then limit the model to validated responses. Sherwin connects this directly to Agent Builder: deterministic nodes give intelligence room to reason while keeping actions inside regulatory, operational, or game-system boundaries.