The hot-swap era for models is over, OpenAI's enterprise product chief says

Olivier Godement sits down with Jacob Efron at OpenAI headquarters as head of product for the company's business products. Last week the lab had shipped GPT-5.1 and GPT-5.1 Codex, and Godement says both were trained on what users actually did with GPT-5. People loved the intelligence, the instruction following, and the way the model streamed. What they did not love was the wait: GPT-5 was strong when it thought for a long time, and sluggish on basic queries.

The design goal for 5.1 was therefore narrow. Keep the intelligence, compress the thinking tokens, and answer much faster. Godement thinks that goal has been met. Users now switch fairly seamlessly between low thinking effort on ordinary questions and higher effort on harder ones, which is the usage pattern the team wanted.

Codex, he says, is having a moment, and it did not surprise the people who live with it. It is the model OpenAI dogfoods most, because every software engineer and researcher has used it. The latest internal stat he cites is that engineers on the team push something like 70% more because of Codex, which has become part of the tool stack rather than a demo. Most of the unlocked work is the expected kind — coding productivity, knowledge lookup, customer support — only more of it.

The domain that has surprised him over the past few months is science. Researchers are using GPT-5 and 5.1 to condense literature and test hypotheses, and for the first time since he joined he feels the company's old goal of accelerating scientific research is actually happening. Mark Chen, OpenAI's chief research officer, worked with a physicist on black holes who handed the model a paper that had just been released, so it was not in the training set, and asked it to reproduce the math. GPT-5 Pro took about 30 minutes. The author told them it was weeks of work for a professional physicist.

Efron puts Andrej Karpathy's recent appearance on Dwarkesh Patel's podcast on the table: the industry is jumping too far, some of the output is slop, and it will take at least a decade before AI meaningfully automates entire jobs. Godement does not pretend agents are easy. Building a really good one is hard, and OpenAI has not reached a world where you spend a day and automate any white-collar role. He does say a few fields already show strong automation, and coding is the one that comes first to mind.

The line he draws is vivid. If he took AI coding tools away from software engineers, there would probably be a riot; people would polish their résumés. That is not the same as fully automating the job of a software engineer, but he thinks there is a line of sight to get there. Customer experience is the other early case. He has been working with T-Mobile on better customer service and says they are starting to hit fairly good quality at meaningful scale. The recipe on top of a strong model is a harness that connects it to tools, a serious evaluation framework, and a human-in-the-loop flywheel. It is a lot of work. His sense is that the amount of work that can be automated reliably over the next year or two will still surprise people.

His next bet is often life sciences. Amgen exists to design and develop new drugs, and that work splits into real R&D — experiments, validation, scientists — and a huge administrative tail. Once a drug recipe is locked, getting it to market can take months and sometimes years of complex regulated documents, committee review, and regulator review. Models are good at aggregating structured and unstructured data and spotting diffs across versions. His hunch is that once regulated industries can version, audit, and permission these systems internally, adoption will jump, and the outcome will be more new medication.

The same pattern shows up in finance. He had recently met a large investment firm whose job is to aggregate and analyze mountains of data in real time and then make judgment calls. That is a model superpower, and he is already seeing adoption inside hedge funds, investment firms, and banks. Heavily regulated paperwork-heavy industries, Efron notes, are almost a perfect large-language-model use case. Godement's point is that the bottleneck is no longer whether the model can read the pile. It is whether the company can put versioning, audit, and permissions around the pile.

Efron asks when OpenAI should sit with a T-Mobile or an Amgen itself, and when the application layer — a Sierra for support, a specialist for pharma — should own the workflow. Godement's lesson from a couple of years of enterprise adoption is the size of the problem. Pick any of those companies, including BNY Mellon, and the internal complexity required to operate at that scale is huge. OpenAI has no illusion it will be the only one building good agents. Part of the mission is to enable third parties, and Apps in ChatGPT, announced at Dev Day, is one early vector: employees already treat ChatGPT as a universal interface and want to do more inside it than OpenAI features alone can cover.

He does not think ChatGPT replaces every tool. A financial analyst will still spend the day in Excel or some purpose-built system. He does expect it to become the first place people check in the morning. Pulse, for him, has become a monumental feature: last night's email, the meeting coming up, the paper that dropped, assembled into something he finds accurate and useful. Simple actions will happen more often inside ChatGPT; deep work still double-clicks out to an IDE or a spreadsheet.

On the research side he says he always runs three clocks. Horizon one is unlocking use cases with current capabilities if you finally have the right harness, the right data, and the right problem. Horizon two is a few months of smart post-training, which is how GPT-5.1 made some experiences faster, more aligned, and more steerable. Horizon three is a fundamental breakthrough in the shape of o1: what you could do if the model could reliably think for 30 minutes. That last one is hard to schedule, but when it lands on a product customers already care about, that is where the magic shows up.

The frontier he most wants for business use cases is continuous learning. He thinks of an agent as a day-one intern with academic knowledge and almost no practical knowledge, because nobody has documented everything they do. Today the relationship is feedback and prompting. Once the model can actually update its weights from human feedback at inference time, he expects another unlock in coding, customer experience, and finance. Efron fills in the image: the intern still shows up each morning, but it would be nicer if they had slept on it and gotten smarter. Godement agrees.

A year ago Efron would have put product-market fit in coding, customer support, and maybe healthcare and legal. He still sees those as dominant, with some interesting voice apps, and asks what else belongs in the "insane product-market fit" bucket. Godement reminds his team that coding, customer experience, and finance are gigantic markets, and that the market for software is larger than software-engineer salaries because there is a shortage of engineers. After the spray-and-pray years after GPT-3, his philosophy is to double down: if you automated tier-one tickets, what does it mean to go further and turn support into a revenue maker? Life sciences and healthcare were not obviously on last year's list; they are on this year's, and some of that is software and change management rather than a new model skill.

Even if research froze for a few years, he says there would still be many years of valuable enterprise adoption. Markets tend to reach a tipping point where not trying starts to look irresponsible, and customer support is already there. A few pioneers take the risk while the thing is shaky and over-scaffolded; they harden it; everyone else follows. That motion, he says, repeats across industries.

Are the harnesses similar across life sciences, support, and coding? Fairly bespoke, which he calls a challenge. People try whatever it takes: one agent, multiple agents, deterministic gates in between. There is not yet a standard agent architecture or runtime adopted across industries, and OpenAI is working on that. His boring explanation for the gap is time. He has been at the company about two and a half years. Year one was keeping up with growth. Year two was sitting down to ask what actually worked. Only now are they converging on which pieces of the stack, if standardized, would meaningfully accelerate adoption.

The standard he can already see is a computer for the agent. Writing code is a more general-purpose capability than software engineering, and models are on an insane progress path at it: write a script, execute it in a shell, get the result back. MCP, he thinks, was a good standard for data and API connections, and the industry will keep normalizing around it. Agent-to-agent communication has not had a true breakthrough that people actually use. He compares it to asking him to work without a laptop: not quite useless, but not far from it. Every improvement in the model's coding ability then has orders of magnitude more impact, because the rest of the harness can be moved onto the vector the models are improving fastest.

Cost is still a limiter, even after a drop Godement calls basically unheard of. Over two or three years OpenAI reduced the cost of a GPT-4-level query by one to two orders of magnitude without sacrificing margin, by compressing model size, using better hardware, and networking GPUs more efficiently. There is still room in every layer of the stack. Coding already pays: multiplying a software engineer's productivity by two can be worth dozens or hundreds of dollars a month. Many other uses are blocked. Models are good at personalization and intent, so why isn't every content homepage infused with a live model? Cost and latency, he says. Dozens of price cuts over two years have each grown volume more than they cut price, which he reads as wild demand still capped by cost. If the destination is a thousand agents running in parallel in the background, the industry has to keep taking cost out of every layer.

Reinforcement fine-tuning is in the zeitgeist, and adoption is still early. Most of the enterprise market is still catching up to GPT-5.1's base capabilities. A few innovators are blocked at the frontier. He worked with an accounting-software firm that needed extremely accurate tax analysis; the model was not quite good enough, and too slow, out of the box. A few dozen high-quality environments and graders lifted it 20 or 30% on their gold eval, which was the difference between not viable and viable. It is heavy work: you have to trust the grader, and a reinforcement-learning job can run for hours or days. OpenAI shipped an RFT API that he believes was the first on the market. He suspects most enterprises will have some RFT use cases, but those will not be the gist of the business. Day-to-day automation should ride the base model.

When developers pick a model, he sorts the decision into three buckets: capabilities and behavior, cost and latency, and vibes — meaning Twitter. The third bucket is getting more important because Twitter is the eval now. Gemini 3 had dropped the day they were talking; the percentages looked good, and Efron still wanted to see what Twitter said. Serious teams do not prematurely optimize for cost. They try to make it work first. Academic benchmarks are a weak signal if you are building a support agent: SWE-bench and MMLU tell you something, but not much. Industry benchmarks are catching up, including Tau-bench for services and GDPval, which OpenAI released to test real-world economic tasks. Startups cannot wait, so they do qualitative testing, and Godement is starting to see people with genuine taste for model nuances, the way good writers have taste they cannot fully put in a framework.

That is why hot-swapping is dying. Efron cites Cognition having to do a ton of net-new work when a Claude Sonnet 4.5-class model landed. Godement calls it model fatigue. For non-trivial use cases, the days of changing one API parameter are basically gone. Idiosyncrasies across providers are getting more distinct: some models respond better to certain instructions, some were post-trained on specific tool signatures and names, some handle long-context recall differently. You have to learn the quirks and retune the prompt and the harness. Even strong startups would rather not do that unless the jump is meaningful. Enterprises, whose actual job is designing drugs or selling phones, want a regular cadence, clear changelogs, and a sense that the next version is three months out. The best teams take the time to battle-test, come in with taste and an open mind, and send specific examples, because OpenAI still controls the post-training and can tweak the next snapshot.

Voice is the other frontier that moved in the last year. GPT-4o, which landed around May or June 2024, was for Godement the second time he felt the AGI after ChatGPT: a model that could express a range of tone and emotion and understand it in a human. He is pretty sure he could no longer distinguish a person from a bot on text. Voice has not crossed the Turing test. Interruptions and cadence are still off. The next job is naturalness: given the same intelligence, you should be equally fine being served by AI or by a person. Deployments are starting in tier-one support calls, for two reasons he cares about. The model is infinitely patient. And multilingual coverage matters more than he had realized. A big U.S. retailer cannot staff every language in the long tail, so it used to make hard trade-offs and lose customers. Those multilingual results, he says, have been strong.

Back on coding, he describes the Codex team as the epitome of a small, talented group singularly focused on a use case, cranking on model, harness, integration, and data. Models are now good at generating and understanding code, but that is only part of a software engineer's job. The rest is being on call, communicating with teammates, scoping changes, making architectural decisions, deprecating APIs. Improving collaboration is the major unlock if AI is going to spread beyond the editor. A more boring blocker is enterprise plumbing: many companies are still stuck on GitHub Copilot v1 because they never finished security and provisioning so agents can actually touch the codebase.

He is starting to see a critical mass of enterprises lean in, provisioning hundreds or thousands of licenses and letting engineers experiment. His periodization is the one that opened the episode: 2024 was the year of coding, 2025 is the year of coding in the enterprise. For 2026 he shrugs and bets on PMs — an automated PM — once models are more reliable at writing and understanding code and get good at collaboration, which together make a more multi-dimensional AI software engineer. Model and harness are getting harder to disentangle. Some of the best agents are trained for a specific harness, which is why he thinks Codex is so good. He talks about them as almost Siamese.

That is also why OpenAI open-sourced the Codex harness: the actual code on GitHub, plus the tool definitions, so you can use the full Codex abilities inside Cursor or any other IDE. He thinks model providers will move from inference APIs to shipping the model, the harness, and maybe a UI — a standard blueprint. The biggest learning of two years building for developers and businesses is that you cannot just drop a new model into an API and expect people to extract its capabilities. Models are beautiful and weird; without a recipe they are hard to leverage. His gut is that enterprises will mostly buy harnesses and solutions. The exception is a use case that is the core business. A retailer should not build its own sales, finance, and IT agents any more than it builds all of its software. People underestimate how much work it takes to reach a high level of quality, and that will be even more true for agents. In the long run he expects convergence to win, after a thousand flowers bloom.

Asked for a cheat sheet from the most sophisticated customers, Godement estimates he has now worked with 200-plus enterprises, including T-Mobile, Amgen, Salesforce, BNY Mellon, and Databricks. Tip one is classic enterprise software: if your data is a mess, you will achieve nothing. He could hand over the most powerful coding agent; if you cannot plug it into the right codebases, identity permissions, and databases, it will not be useful. A lot of the work is explaining how to structure data, stand up services when there is no API, decide whether to use MCP, authenticate requests, and log them. It is data engineering, and it is step one.

Step two is teaching teams how to evaluate. Vibe evals are necessary to a degree, but a production-grade use case needs rigor, and that does not come naturally unless a team has done it before. OpenAI talks a lot about the people who train models. It also has a crack team that evaluates models full-time, and he thinks they are probably equally important. Spend the time documenting golden sets, documenting standard operating procedures, and building evaluations you can actually hill-climb. The most common mistake is too little or too much: only five eval sets, then the first real customer arrives completely out of distribution and you learn nothing. His biggest finding inside companies is that most of the knowledge is in people's brains. You are lucky if 20 or 30% of the procedures are written down in Confluence. The rest is "Sarah and Mark know this — you should talk to them." Building evals is therefore less about converting text and more about finding the right people, then iterating: ship, notice what is wrong, ask John why, and add the case.

Step three is change management, which he cannot emphasize enough. The technology is new to everyone, sometimes unnerving in how powerful it is and sometimes frustrating in how unintuitive it is. Taking the time to explain how it works is not optional. On a newer surface, Sora's API, he is already seeing energy in ads and content generation and inside studios, where showing someone in 30 seconds the picture in your head is useful for brainstorming. Video generation is still early, expensive, and slow, but he can see how it will change those jobs.

In the quickfire he puts science and drug discovery as the most underhyped bet in AI. Software people do not reach for it naturally, but it is the substrate of progress; even a 5% faster rate of discovery would have enormous knock-on effects. Doing it well means a lab, experimental data, and people at the intersection of language models and the field — hard, and probably the highest compounding benefit. The thing he changed his mind on in the last year is that the model is everything. Harnesses are getting extremely important, the two are hard to disentangle, and enterprise adoption also needs standard infrastructure to present the right data at the right time. Looking back, the world has been forgiving because OpenAI is often first; 2023's definition of an agent was too early and some APIs had to be sunset, and some audio bets did not pan out. He has not played with Gemini 3 yet. When he tests a new model he has three personal probes: dump a lot of context and ask private questions about style and tone; generate an app from scratch and look at the frontend; and dump hundreds of thousands of tokens, then ask a hard question about the second-to-last one. He is on Twitter, and he says he responds to everyone.