OpenAI vs Anthropic Revenue 2026: Why the $25B–$19B Gap Is Misread

What these numbers are

Per Axios (Apr 13, 2026), Anthropic's annualized revenue run-rate passed $30B, up from roughly $19B in early March and about $9B at the end of 2025. The title unpacks that same early-March comparison — the $25B-vs-$19B gap that circulated in coverage before the $30B figure landed. On the slope: Anthropic climbed from roughly $9B at the end of 2025 to roughly $19B by early March, and later public reporting put the run-rate above $30B. The same reporting puts OpenAI as the closest comp at roughly $25B annualized. Both figures are self-reported to press. Neither is audited. Neither is trailing-twelve-months. And they are not built the same way.

One caveat on the baseline. Public reporting gives a $9B end-2025 figure and a $19B early-March figure — that $9B → $19B path is the conservative anchor for the slope. Model from the public $9B → $19B path, and label the $19B as an unaudited, self-reported figure — not a GAAP line you can lean on.

The "gap" — $25B minus $19B — treats those two numbers as one currency. They aren't. OpenAI's mix leans heavily on subscription seats. Anthropic's skews toward pay-per-token consumption. Subtracting one from the other is like subtracting a bond yield from a stock price and calling the result a spread.

Everything below is why that matters, and what would actually signal Anthropic closing the gap.

Run-rate is a level, not a pace

Annualized revenue here means one thing: take the most recent month's recurring or consumption revenue and multiply by twelve. That is the whole formula. It is not GAAP-recognized revenue.

Now watch what a $9B → $19B move does mechanically. When AI ARR figures jump double digits in a fortnight, the explanation is almost always re-annualization, not compounding growth. One large enterprise contract activates mid-month, or a price change or a rate-limit lift lands, the latest month jumps — and ×12 amplifies the jump. What you are seeing is the level re-basing, not the business compounding at that rate every fortnight.

You cannot trend that forward. Multiplying it out — "it doubled in months, so it will keep doubling" — is the exact error the headline invites. A re-annualization off a lumpy consumption base is supposed to move in steps. The step is the news. The slope is imaginary.

Whenever you see an "annualized" AI revenue figure, write next to it: latest month × 12, unaudited, self-reported. That one annotation immunizes you against most of the coverage.

The ruler problem: seats vs tokens

Seat revenue and token revenue have opposite stability profiles, and the two companies lean on different ones.

OpenAI carries a large ChatGPT consumer, Team, and Enterprise seat base. Someone pays for the login whether or not they use it heavily. Monthly churn is slow — cancelling means re-onboarding and re-procurement. Predictable.

Anthropic is disproportionately consumption. Claude API revenue is structurally different from a seat subscription: it scales with tokens, not logins. Revenue equals requests times tokens times price, and every one of those is a dial the customer holds.

So when both are labeled "annualized revenue" and subtracted, you are comparing a recurring-seat number to a metered-usage number as if a dollar of each were equally durable. It isn't. One is a subscription. The other is a utility bill that can drop overnight.

Why token revenue swings so hard

In seat SaaS, usage and revenue are decoupled. In token and inference billing, they are the same thing, live.

A model swap is a config change. Workloads route through a gateway — LiteLLM, or an internal router. A customer can move 40% of spend to a cheaper or better model in an afternoon. No annual contract holds the usage in place — only the floor of a minimum-commit, if one exists.

And the bill scales with the shape of the request, not just the count. This is where long context — the thing everyone is excited about — turns nasty. Attention cost grows super-linearly with sequence length. A customer moving from 8K to 200K-token prompts inflates the token count and the per-token serving cost at the same time.

Which sets up the trap in the next section.

The margin trap

"Revenue doubled" tells you nothing about whether gross profit did.

If part of that revenue surge is customers sending much bigger, longer-context, interactive requests, you can get revenue up while contribution margin falls. More dollars, thinner dollars. The top line and the unit economics can point in opposite directions, and an annualized run-rate reports only the top line.

So a surge could be a more durable business — or it could be the same customers running more expensive requests. From the outside you cannot tell. That ambiguity is the whole point.

Where the margin actually lives

The cost to serve a token is dominated by GPU-seconds, and three levers move it hard.

Prompt caching, or KV-cache reuse. If a customer sends the same long system prompt every call, the provider can cache its key-value tensors and skip recomputing them. A cache hit is cheap. A cache miss — fresh context every call — is expensive. Anthropic exposes explicit prompt caching precisely because it moves margin materially.

Batching. Packing many requests into one GPU forward pass raises utilization and cuts per-token cost. Bursty, latency-sensitive traffic — like interactive coding agents — batches poorly.

Speculative decoding. A small draft model proposes tokens the big model verifies in parallel, cutting latency and cost on predictable outputs.

The consequence, using round illustrative numbers: the same $1 of API revenue might be high-margin when cached, batched, and short-output — or negative-margin when cold, long-context, interactive, and low-batch. Aggregate revenue growth tells you nothing about which mix you got.

If you run any nontrivial API spend, open your provider's prompt-caching docs and check whether your repeated system prompt is actually being cached. A stable long prefix that isn't cached is the single most common way teams pay full price for tokens they could reuse.

Capacity caps the growth being celebrated

Foundation-model APIs run on reserved GPU capacity. Providers typically operate on multi-year GPU reservations from AWS, GCP, Azure, or owned datacenters — the exact terms aren't public, but the general shape is: capacity is committed ahead of demand, not bought on the spot. Revenue can spike faster than tokens can be served.

When demand surges past served capacity, the provider protects its latency SLAs with rate limits and tier throttles. So a single demand spike can, at once, inflate the run-rate off contracts already being served and be unservable going forward until more capacity lands.

The number celebrating the demand and the constraint suppressing it are the same event. That is a real reason to doubt naive extrapolation — not a bearish opinion, a physical one.

Where the demand physically comes from

Buyers route by task, not loyalty.

Agentic coding — refactoring, long-horizon tool use — concentrates on Claude. Claude Code, Anthropic's own agentic coding tool, runs on Claude by default and out of the box; you can also point it at other backends via third-party proxies, OpenRouter, local models, or a compatible API. Cursor, a separate agentic coding tool, supports multiple models including Claude. Consumer and general reasoning, plus a massive free-to-paid consumer funnel, concentrate on OpenAI.

Anthropic's surge is concentrated in the coding-agent power-user segment: a relatively small number of customers burning tokens at per-user rates no seat model would predict. High growth, yes — and concentrated, and re-routable. A handful of large accounts, each holding a router that can move spend.

Booking event or consumption event?

There are two clean readings of the spike, and you cannot distinguish them from outside.

One: a booking event — a large committed-spend deal or a rate change re-basing the run-rate. Structural, but a one-time step. Two: a consumption event — coding agents scaling seats times tokens organically. Real demand, even if margin-thin.

It is probably both. The honest position is that it is uninterpretable without disclosure neither company gives.

One more wrinkle worth flagging: a large chunk of Anthropic's usage routes through AWS Bedrock and GCP, and that raises a rev-rec question the reported figures don't settle. Is marketplace-mediated revenue booked gross or net? Rev-rec there is murky, and committed-spend deals are usually minimum-commit floors, not usage guarantees — a $100M three-year commit can be a $33M-per-year floor the customer is nowhere near burning. Read that as editorial synthesis, not a reported fact about either company's books.

Is the moat real? It depends on the task

Switching cost at the API layer is not one number.

For generic chat and completion, it is near zero. Prompt-portability plus a decent eval harness migrates a workload in a sprint. "Catching up" here just means who won last quarter's routing decision.

For tuned agentic pipelines, it is meaningfully higher. Once you have tuned tool-use scaffolding, prompt-cached your system prompts, and calibrated to a model's specific failure modes, migration is not a config flip — you re-run the whole eval suite and re-tune the agents.

So the moat exists, but mostly in the exact segment Anthropic happens to lead. Stickiness is task-dependent: low for chat, real for agents.

What would actually signal the gap closing

Not "annualized revenue up." Watch instead for the things that survive re-annualization.

Trailing-twelve-months revenue, or at least a multi-month recurring line, rather than a single month times twelve. Net revenue retention on token accounts — is existing spend expanding or churning? Gross margin trend, so you know the growth isn't long-context revenue at a loss. Concentration — what share of the run-rate is the top handful of coding accounts? And the gross-versus-net treatment of Bedrock and GCP-routed revenue.

Until those surface, the "arch-rivals, one gap" frame is telling you two structurally different businesses are running the same race. They aren't.

FAQ — the questions the headline doesn't answer

What does Anthropic revenue growth signal? The surge in Anthropic's annualized run-rate signals genuine enterprise adoption of Claude in agentic coding workflows — but it also reflects re-annualization mechanics that amplify lumpy token consumption. The underlying demand is real; the slope implied by the headline is not.

What does Anthropic revenue growth signal for the API market? A real token-consumption boom driven by coding agents burning tokens at rates no seat model predicts. But it also flags concentration risk: the fastest-growing, most defensible revenue sits with a small number of large accounts, each holding a router that can move spend overnight.

Why can't you subtract OpenAI's $25B from Anthropic's $19B and call it a gap? Because one number is mostly recurring seat revenue and the other is mostly metered token consumption, and a dollar of each has opposite durability. It's the ruler problem — you're subtracting a subscription from a utility bill.

What is annualized revenue run rate and how does it differ from reported revenue? Run rate is the latest month's revenue multiplied by twelve — a level, not a trailing figure. It's unaudited and not GAAP-recognized, so it moves in steps whenever a big contract or price change re-bases the latest month.

Is Claude API revenue growing as fast as the headline suggests? The demand is real, but the headline slope isn't. When AI ARR figures jump double digits in a fortnight, that's re-annualization off a lumpy consumption base — a one-time re-basing, not a rate the business compounds at every two weeks.

Who wins the AI model business competition? The AI model business competition is usually framed as a revenue race, but the more useful frame is recurring vs. consumption mix. Whoever holds the durable, high-margin, low-concentration revenue wins — and a headline run-rate can't tell you who that is.

What would actually signal Anthropic closing the gap with OpenAI? Not "annualized revenue up." Trailing-twelve-months revenue, net revenue retention on token accounts, a rising gross-margin trend, lower account concentration, and clarity on gross-vs-net treatment of Bedrock and GCP-routed revenue. Those survive re-annualization; a single month times twelve doesn't.

Investor Checklist — three things to do tonight

Re-annotate the headline. Write beside any AI revenue figure: latest month × 12, unaudited, self-reported — then stop extrapolating it.

If you spend on an API, open the prompt-caching docs and confirm your repeated long prefix is actually cached. An uncached stable system prompt is money left on the table.

Change the question you ask. Instead of "who's ahead on revenue," ask "seat revenue or token revenue, and at what gross margin." That is the split that tells you which lead is durable.

One open question is worth sitting with. If Anthropic's growth really is the coding-agent segment, its most defensible revenue is also its most concentrated and most compute-hungry. Is that a moat, or a single point of failure? If you are routing real spend across these models, which way is it breaking for you?