OpenAI published an essay this week arguing that cheaper, more capable models create a self-reinforcing loop: lower costs pull in more users, more users generate revenue and feedback, and that funds the next round of infrastructure and research. It is worth reading the piece for what it actually is: the company that needs the most capital in the industry making the public case for why that capital keeps paying off. That framing does not make the argument wrong. It does mean each link in the chain deserves separate scrutiny rather than a nod at the whole picture.
The cost side holds up because OpenAI backs it with numbers. GPT-5.6 Luna’s price dropped 80 percent: OpenAI now charges $0.20 for every million tokens a request sends in, and $1.20 for every million tokens it generates in response. GPT-5.6 Terra’s price fell by a smaller 20 percent, landing at $2 and $12 for those same two token counts. A new Fast mode for GPT-5.6 Sol turns in results 2.5 times quicker than the standard setting, for double the cost, with intelligence unchanged. Those are checkable commitments, not vision language, and together they describe a real drop in what a unit of intelligence costs to buy.
The adoption claim is where the essay starts asking for trust rather than showing proof. By OpenAI’s own count, the active user base has crossed one billion, and the business tally has passed two million. The company layers a growth claim on top of that scale: users who stick around for six months are sending about 50 percent more messages a day than when they started, and the range of tasks they bring to the product has roughly doubled over that same stretch. Those figures describe usage climbing over time among people already on the platform. They do not establish that this week’s price cuts, specifically, are what pulls in the next cohort of users. The company asserts the causal link between cheaper tokens and broader adoption; it shows only that adoption has been rising anyway.
The middle of the loop, where adoption is supposed to convert into revenue and reinvestment, gets no numbers at all. No revenue figure appears anywhere in the piece, and neither does a capital commitment tied to a specific facility, region, or timeline. Given that OpenAI has spent much of the past year discussing infrastructure deals worth hundreds of billions of dollars in other venues, the absence of a dollar figure here is notable. The reader is asked to accept “revenue funds the next generation of infrastructure” as a premise rather than a demonstrated result.
The efficiency claims are the strongest part of the piece, and they come with the most specific evidence. GPT-5.6 Sol reportedly helped optimize the software that serves OpenAI’s models: the pipeline now costs 20 percent less to run from request to response, and a separate change to speculative decoding made token generation more than 15 percent more efficient. OpenAI also points to how GPT-5.6 Sol performed on ARC-AGI-3, a public benchmark: its score had been 13.3 percent, then reached 38.3 percent after a round of changes, needing six times fewer output tokens to get there. The model itself did not change. What changed was the software layer built around it, the settings that control how it retains reasoning and manages context between steps. That is a concrete, sourced example of software and system design doing work that infrastructure spending alone would not. It is also one benchmark on one internal system, not proof that every dollar of new data center capacity converts into gains of that size.
One more data point complicates the “asking to doing” narrative OpenAI leans on: agentic work through Codex now makes up 99.8 percent of weekly output tokens across the company, with the finance team named as an example of a group that has adopted agentic tools as a primary workflow. That is a real shift in how OpenAI’s own employees use its products, but it describes internal usage, not the enterprise market broadly.
Put together, the essay proves the price side of its own argument and offers a genuinely interesting efficiency case study, while leaving the revenue and reinvestment middle of the loop as assertion. Anyone weighing OpenAI’s infrastructure buildout as an investment thesis should treat the pricing and efficiency numbers as real data points and treat the adoption-to-revenue link as the part still awaiting evidence, likely in a future earnings disclosure or funding round rather than a blog post.
OpenAI published the essay, titled “Building abundant intelligence,” on openai.com; the post carries no dateline, though it references pricing changes the company rolled out the prior day.
