OpenAI's Agents API Sells What the AI Trade Runs On: Inference Volume

On September 10, OpenAI put its Agents API into public beta — the company's managed service for building "long-running agents." An agent built on it doesn't answer a single question; it works through a multi-step task, calling tools, spawning subagents, and compacting its own context until the job is done. A developer defines the whole thing in one API call — the task, the model, the tools — and OpenAI plays host, running the harness and sandbox on the same machinery that powers Codex and ChatGPT for Work.

For anyone holding the AI trade, the detail that matters is on the billing line. OpenAI charges no separate fee for the API itself: developers pay for the tokens the agent consumes. Strip away the product language and what OpenAI is selling is metered computation, and it gets paid whenever an agent spends.

That billing structure is the whole ballgame, because agents are spectacularly bad for your token budget. Gartner's 2026 analysis of agentic workloads puts the multiple at five to thirty times the tokens of a standard chatbot exchange — resolving one support ticket can burn as much as a few dozen ordinary conversations. Every tool call, every subagent, every context compaction is metered output. So the Agents API doesn't just enable agents; it licenses them to be compute-hungry. That is a small change in vocabulary and a large change in what actually drives demand.

The launch story is not the product. The launch story is an inference appetite.

None of this adds a dollar to OpenAI's value on its own, because OpenAI is private — you can't buy the thing that benefits most directly. What it does is tip the broader AI investment story from "models keep improving" toward "how much inference compute gets burned," and that lands in public markets through the stack that serves it: Nvidia's accelerators, Microsoft's Azure, which hosts OpenAI, and the networking and data-center layer underneath. Nvidia has already run up on this theme; and on the software side, analysts who parsed Microsoft's disclosures put OpenAI at around 70% of Microsoft's AI revenue. For the owners of those names, the Agents API is a bet that agentic usage actually shows up as paid volume, not a demo.

The bill is the constraint, and it's OpenAI's own bill

Here is where the anti-PR check earns its keep. OpenAI frames agents as a growth story, and the revenue is real — roughly $40 billion annualized by mid-2026, up from about $20 billion at the end of 2025, with enterprise more than half the mix. But the company is also a machine that spends more than it earns. Independent analyses put it at something like $3.30 spent for every $1 of revenue, and project a compute expansion, roughly $750 billion, it cannot fully cover out of its own operations; a leaked-audit narrative renewed that worry in early September. Fast-growing and unprofitable is the startup norm — but at this scale, the gap between what agents promise to earn and what the compute to serve them costs is not a footnote.

OpenAI is trying to close that gap on the supply side, and this is where a launch that looks like software actually points at silicon and economics. The company's first custom inference chip, Jalapeño, is shipping first results with higher throughput per kilowatt and lower token latency than the commercial systems it's compared against. It is also attacking cost on the model side: GPT-5.6 Sol reportedly reaches a top coding-benchmark score using roughly 54% fewer output tokens than a leading rival. OpenAI itself frames all of this as a Jevons paradox — the cheaper each useful unit of intelligence gets, the more tasks become economical, and the more total consumption grows.

It is a self-serving argument, sure. It is also the mechanism. Unit cost falls, volume rises, and the entire bet rests on which one wins out.

What to actually watch

So the Agents API pulls two levers at once, and they pull against each other. On the demand side, it converts agents from a demo into metered, long-running workloads that could multiply inference consumption — a tailwind for the chip and cloud names that retail investors can actually own. On the cost side, OpenAI's whole compute strategy is engineered to shrink the price per unit of that volume, and its balance sheet is straining to fund the compute in the first place.

The observable that settles this is not the launch, and not OpenAI's slide deck. It is whether agent token consumption materializes in paid production — in API spend, in Azure and cloud demand, in the revenue disclosures that flow to Microsoft — and whether unit costs fall fast enough to keep that volume profitable rather than just bigger losses. The beta customer numbers OpenAI chose to publish (one partner claiming a 60% cut in cost per case) are marketing; the only real data will be paid usage over the next few quarters.

When the mechanism and its economics point in a clear direction, you stop there. Here they don't fully: the mechanism — agents multiply tokens — is clear, and the economics — who profits from that volume — is precisely what's unresolved. The Agents API removes the engineering doubt about whether an agent can run a long workflow. It does nothing to remove the doubt about who ends up paying for the compute.