Anthropic disclosed an annualized revenue run rate of roughly thirty billion dollars in April 2026, against an OpenAI figure of roughly twenty-four. It is the first time the public chart of the two leading model labs has gone in this direction. The trajectory under the headline is the part that took everyone off guard. Anthropic was at one billion in ARR in January 2025, around nine billion at the end of 2025, and thirty billion fifteen months after the first reading. Several people we trust on early reads of these numbers initially assumed they were typos. They were not.
The number is real. It is also, depending on how you count, somewhere between one and two different products being added together. OpenAI’s internal pushback, reported from a memo by CRO Denise Dresser, argues the apples-to-apples figure is closer to twenty-two billion once accounting treatment is normalized. Whether the gap is eight billion or four does not actually move the more interesting question, which is what each company’s revenue is made of and why those compositions have diverged so sharply.
Where the Anthropic number is actually coming from
Roughly eighty percent of the figure is enterprise and API revenue. Anthropic reports more than three hundred thousand business customers and over a thousand accounts spending more than a million dollars a year on Claude. The million-dollar count reportedly doubled in under two months. The ARR composition is heavy on programmatic API consumption that grows in a usage-driven, multi-year contract pattern, which is a different shape than the consumer subscription pattern that grew the previous generation of AI revenue.
Inside that mix, Claude Code is the line that matters most for what the next twelve months look like. It crossed one billion in ARR in November 2025 and reached two and a half billion by February 2026. A code product hitting that run rate inside a single quarter is unusual. It is a category where the buyer is the engineering team rather than central IT, where the trial-to-paid loop is short, and where usage scales with how much code the team is shipping rather than how many seats it has. That category was not sized at multiple billions of dollars in any analyst model we read in 2024.
The other piece worth holding next to the revenue figure is the reported training cost, which is somewhere around a quarter of OpenAI’s per-model spend. We do not have audited numbers on either side, but the directional claim has held up across several independent reports. If the ratio is even close, the implication is that Anthropic is converting research dollars into model capability at a different rate than the company it just passed. That advantage transfers downstream into price elasticity on the API.
Where the OpenAI number is actually coming from
The OpenAI figure is heavier on consumer surface revenue. ChatGPT subscriptions, ChatGPT Enterprise, and a serious but proportionally smaller API book combine into the published number. The product surfaces are different enough that they grow on different curves. Consumer revenue compounds with conversion, churn, and pricing per seat. API revenue compounds with token volume per customer per month, which is a different function and tends to be less elastic to plan changes.
The CRO memo arguing the comparison is unfair makes a real accounting point. Some of the difference between thirty billion and twenty-four is treatment of multi-year contract value, deferred revenue recognition, and what counts as recurring inside an annualized run-rate framing. We would not dismiss that point. We would also note that even at twenty-two billion, the more durable observation is not who is in front this quarter. It is that the two companies are no longer growing the same product through the same channel to the same buyer.
The substrate question
Meta and Google ran into the same kind of divergence in 2026 on the ad side. Two companies on the same leaderboard, different products underneath, different growth functions, and a chart that obscured the structural fact by averaging across them. The Anthropic and OpenAI numbers sit on the same shape. Anthropic looks like a B2B model platform with a developer-led sales motion and an unusually strong code product attached. OpenAI looks like a consumer AI product company with a serious enterprise side bolted on. Both are coherent businesses. They have different ceilings, different cost structures, and different ways the next dollar of revenue is earned.
The training-cost gap, if it is real, makes the substrate difference more durable rather than less. A model lab that trains for a quarter of the cost has more shots per year on capability, more room to compete on price without compressing margin, and a different set of choices about which workloads to subsidize. The same dollar of revenue does not buy the same things on the two sides of the chart. That is the actual structural point in the April release, and it is the one that is hardest to read off the headline number.
What changes for the people building on top
Three things move when the leaderboard tilts this way, and the third is the one that takes longest to show up.
The first is that model choice is now a genuine choice, not a default. For most of 2024 and into early 2025, the planning conversation in a typical engineering team treated OpenAI as the baseline and other providers as either ablation experiments or cost-driven swaps. That defaultness is gone. A team writing down its model strategy in mid-2026 has to argue from capability, price, and roadmap, not from market share. The number of teams we see running multi-provider routers has grown roughly with that shift.
The second is that pricing power on bulk inference is going to keep compressing. Two suppliers competing for the same enterprise budget, both with credible roadmaps and serious training-efficiency stories, do not maintain premium pricing on commodity tokens for long. The premium will keep moving up the stack, into the higher-reasoning models and the agentic tooling around them, while the floor on simple completion drops. We have already seen that in 2026’s per-token pricing on the mid-tier models on both sides.
The third change is the one most teams underweight. Switching costs on these models are technically low and practically high. The technical surface is mostly compatible. The practical surface, prompts tuned over months, evals built around one model’s failure modes, tool-calling conventions inside a particular SDK, retry logic shaped to one provider’s rate limits, takes real engineering time to migrate. A team that runs everything on a single provider gets simplicity now and loses bargaining position later. That tradeoff is the one we think will look most different in retrospect, once the second supplier is genuinely interchangeable on the workloads that matter.
What the chart is and is not predicting
A four billion dollar lead on a base above twenty billion is a single quarter of execution. The 2027 reading could put OpenAI back ahead, especially if the next ChatGPT release pulls a meaningful step forward on consumer engagement, or if the API book accelerates faster than the current trend implies. We would not bet on this being a permanent reversal. We would bet on it being the first of several reversals across the next few years, and a sign that the chart is now genuinely competitive rather than the one-horse race it appeared to be in 2024.
The more durable point is not who is in front. It is that the consumer-product theory of AI revenue and the API-platform theory of AI revenue have separated far enough that they are now telling different stories. The interesting numbers to watch are not next quarter’s gross. They are Anthropic’s revenue per enterprise account, especially in the long-tail of accounts that are not yet at the million-dollar mark, and the consumer-to-enterprise ratio inside OpenAI’s mix. Those two numbers will say more about which substrate compounds faster than any leaderboard headline will.