It’s been a little over a month since Sam Altman admitted something you don’t often hear from the CEO of the company that helped kick off the generative AI boom.
“We did not have our best last 12 months ever, which is mostly my fault.”
Then he raised the stakes.
Altman said OpenAI is about to have its best 12 months to date.
A month later, that promise is worth looking at more closely because OpenAI isn’t simply trying to release a better model.
It’s spending enormous amounts of money on compute, pushing deeper into AI agents and products, competing with increasingly capable rivals, and making bets on infrastructure that could shape how the company operates for years.
So what changed?
And more importantly, what exactly is Altman betting on to turn a year he largely blames himself for into OpenAI’s best one yet?
Table of Contents
The Year Altman Says He Got Wrong
Altman never explained exactly what went wrong. He didn’t point to a failed model, a revenue miss, or a single decision that derailed OpenAI. He simply acknowledged that the company had fallen short of where he wanted it to be.
But there were plenty of signs that OpenAI was no longer operating with the same distance from its competitors.
The model race was no longer OpenAI’s alone. Earlier GPT releases repeatedly reset expectations for what AI could do. But as competitors caught up, every new OpenAI model faced a much higher bar. Developers now had more credible alternatives, while some power users were voicing frustration over incremental updates, restrictions and product limitations.
That doesn’t prove a mass exodus from OpenAI. It does show that developer loyalty could no longer be taken for granted.
OpenAI wasn’t suddenly out of ideas. It was operating in a market where models were becoming harder to differentiate and competitors were closing the gap.
The question now is what OpenAI is willing to change to make the next year different.
The First Sign That OpenAI Is Changing Course
The first real clue may have come not from a new model, but from its price tag.
On August 21, OpenAI cut the API price of GPT-5.6 Sol by more than 20% for the next three months. Input pricing fell from $5 to $4 per million tokens, while output dropped from $30 to $20. The reduced pricing also applies to eligible credits for Codex and ChatGPT Work, while consumer subscription prices remain unchanged.
On its own, a price cut isn’t unusual in AI. Models are getting cheaper across the industry as competition intensifies.
But the timing is interesting.
OpenAI is making its frontier intelligence cheaper at the moment when the company needs developers to keep building on its models and rivals like Anthropic and Google are making that choice harder.
The shift may be this: model intelligence is becoming harder to use as a moat. Price, distribution and the products built around that intelligence matter more.
And OpenAI appears to be leaning into all three.
The price cut is only temporary, but it gives us the first concrete sign of what that new strategy might look like.
What Is OpenAI Betting On?
The price cut is only one clue.
OpenAI is betting that increasingly capable AI will become something people use to do work, not just something they chat with. That means developers building agents, coding systems running in the background, and businesses relying on AI for longer, more complex workflows.
That changes what OpenAI needs to win.
Having the smartest model still matters, but it is no longer enough. OpenAI needs developers to build on its models, users to rely on its products, and those systems to consume enough AI that OpenAI can turn its enormous infrastructure investment into a durable business.
That’s where products such as Codex become important.
If AI moves from answering a prompt to completing an entire task, writing and testing code, working through documents, running multiple steps and continuing without constant human input, usage can increase a lot. A developer isn’t just sending a few prompts anymore. They are effectively giving an AI system a job.
And that makes the economics of inference much more important.
A cheaper frontier model can encourage more experimentation. Better agentic tools can create more reasons to use it. More usage can, in turn, strengthen the ecosystem built around OpenAI.
But there is another side to this bet.
OpenAI has to keep spending heavily on the infrastructure required to provide all that intelligence. The company therefore isn’t just betting that its models will become better but more towards AI usage will grow fast enough to justify the enormous cost of building and operating the systems underneath it.
And that’s where the stakes get much higher.
Because OpenAI is trying to make itself one of the companies that controls the infrastructure, products and workflows through which the next phase of AI gets used.
Also Read: Cursor Origin Doesn’t Want to Replace GitHub. (Yet.)
But OpenAI Isn’t Alone in This Bet
The problem is that the market isn’t empty.
Anthropic is pushing Claude deeper into coding and enterprise workflows. Google has something OpenAI cannot easily replicate: a massive ecosystem through which Gemini can reach users. Meanwhile, Chinese AI companies are adding pressure by competing aggressively on price.
That makes the race very different from the one OpenAI dominated a few years ago.
Having a model that simply felt better was once enough to pull users and developers toward it. Now, developers can compare capability, reliability, tooling and cost across several credible alternatives.
OpenAI therefore has to do more than stay at the frontier. It needs its models to be affordable enough, useful enough and deeply integrated enough that developers and businesses have a reason to keep building around them.
And staying in that race requires something far more expensive than a model release: compute.
The Most Expensive Part of the Bet
AI is getting cheaper to use, but more expensive to build. OpenAI needs enormous amounts of compute to train its next generation of models and then keep those models running as usage grows. That is why its strategy increasingly extends beyond the model itself and into the infrastructure underneath it.
The Stargate project is the clearest example. OpenAI and its partners announced plans to invest up to $500 billion in AI infrastructure over four years, with the goal of securing the computing capacity needed for OpenAI’s future systems.
And the scale of that infrastructure race is still growing. Nvidia has now committed more than $100 billion in credit support for a new OpenAI-linked data center project in Ohio that is expected to eventually reach 8 gigawatts of capacity.
That tells us something important about Altman’s “best 12 months” claim.
OpenAI isn’t betting that one breakthrough model will suddenly solve its problems. It is betting on a much larger cycle: better models -> more usage -> more compute -> more capable AI -> even more usage.
But there is another piece to the strategy that could put OpenAI even closer to users.
The company is now building a family of AI devices. OpenAI President Greg Brockman recently confirmed that hardware is coming, while reports indicate its first device could be a screenless AI speaker designed around interacting with ChatGPT throughout the day.
That may sound like a completely different business.
It isn’t.
If OpenAI can move from being an app people open to becoming an AI they interact with throughout the day, it gains something incredibly valuable: a direct relationship with the user.
And suddenly, the pieces start fitting together.
OpenAI wants the models. It wants the developers building on them. It wants the compute to run them at enormous scale. And now it appears to want more control over the devices through which people interact with that intelligence.
That’s a much bigger gamble than simply releasing a better GPT.
And it also explains why the next 12 months matter so much: all of these bets have to start working together.
Also Read: Open Source AI Coding Agents That Don’t Need a Subscription
So, Can This Actually Work?
That is the harder question.
OpenAI’s strategy is clear enough: make intelligence cheaper, push it deeper into real workflows, attract more developers, and build the infrastructure needed to support that growth. The problem is that competitors are making similarly serious bets. Anthropic, for example, is also expanding its compute capacity as enterprise demand grows.
OpenAI therefore doesn’t need to win every model comparison. It needs the pieces to work together, more developers, more usage, deeper workflows and enough compute to make the economics work.
If cheaper inference drives usage and agents turn that usage into long-term workflows, the strategy starts reinforcing itself. If competitors capture those developers first, it becomes a very expensive race.
And that’s ultimately what Altman is betting on: An ecosystem that becomes difficult to replace.
The Next 12 Months of OpenAI
A month ago, Altman said the next 12 months could be OpenAI’s best yet.
The company now has a lot riding on that prediction. Cheaper models, massive infrastructure spending, agentic products and new hardware are not independent bets. They are pieces of the same attempt to keep OpenAI at the center of the AI ecosystem.
Whether they fit together as planned will determine whether Altman’s confidence was foresight or just another ambitious promise.
