OpenAI Just Shipped the Codex Harness Behind One API Call. It Priced the Orchestration Layer at Zero.

OpenAI shipped its Agents API in public beta on Thursday, September 10, 2026. The announcement is short and the sentence that decides the read is one line: the API exposes the same harness and infrastructure that runs Codex, at no additional fee above the usage cost of the underlying models. Subagents, MCP, automatic context compaction, tool search, artifact generation, and a choice of managed sandbox or one of nine partner sandboxes. The orchestration layer, which TensorFeed argued last week was the actual product, just got a public price of zero.

Five days ago Anthropic published the Prove2Me receipt on Fermat. Four days ago OpenAI published the Navier-Stokes swarm receipt. Yesterday OpenAI turned the harness behind the Navier-Stokes run into a shipping API. That is a nine-day cadence from research artifact to product line, and it is the answer to a question we left open on Tuesday: what does OpenAI do with the swarm now that it has proven it works at frontier scale? The answer is put it behind the API and let developers pay for the tokens.

What Shipped, In One Table

Read the right-hand column top to bottom and you are looking at a partial list of the agent-tooling market as it was Wednesday. Every row is a category where at least one venture-funded company was billing seats or transactions for something OpenAI now offers in the default API loop. The Agents API does not make those companies unviable overnight. It moves the price floor for what a bare-metal agent runtime costs to a number that reads zero on the invoice.

The Sandbox List Is the Real Announcement

Nine sandbox partners on day one, in the order the launch materials name them: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel. Six are startups or serverless-compute vendors we have covered on the agent-payments beat, three are hyperscalers or infrastructure incumbents. The partner list matters more than the feature list, because the sandbox is the piece of an agent workload that actually spends money.

Here is the shape of the bill. On a typical long-horizon coding session, the model does a few hundred milliseconds of inference, the tool call runs for several seconds to several minutes inside a container that has to be provisioned, mounted, warmed, and eventually torn down. The sandbox is where the file system lives, where the git checkout sits, where the language server runs, where every intermediate build artifact accumulates. On a Cursor-style workload the sandbox side is roughly 10 to 30 percent of the total run cost today, and on longer research runs it climbs to 50 percent or more because the container is up for the full wall clock.

OpenAI does not want to sell that. Running a fleet of general-purpose Linux containers across nine geographies is a data-center-operator business with data-center-operator margins, and OpenAI has been trying to move up the stack since before Codex existed. Handing sandbox to Cloudflare, Modal, Vercel, and Oracle is the exit from a line item that would otherwise dilute the API margin, dressed up as a partner ecosystem. The revenue-share arithmetic is not disclosed and probably will not be, but the setup is familiar: OpenAI owns the harness and the model, the sandbox vendor owns the compute, and the developer pays both.

What It Does to Prove2Me

Anthropic has a harness that survived a formal proof of Fermat in 11 days on 6 billion output tokens. It is called Prove2Me, it is roughly 200 lines of Python plus a scheduler, and it lives inside Anthropic Research as a working artifact with a paper attached. It is not a product. There is no pricing page, no SDK version number, no partner list, no managed sandbox, and no MCP wiring beyond what the researchers needed for the run. Every question a developer would ask about using it in production has the same answer today, which is that Anthropic has not shipped it.

The pressure that puts on Anthropic is uncomfortable and specific. If Prove2Me is the canonical example of the harness thesis, and the harness thesis is a real product category, then the lab that authored the example does not currently sell the category and the lab that shipped the swarm receipt on the same category four days later does. Anthropic can respond three ways. It can ship Prove2Me as a product before quarter end, which requires SDK, docs, sandbox partners, and a pricing decision the API team has not yet made public. It can announce a competing agent runtime that reuses the Claude Code orchestrator with the pieces that Claude Code hides from third parties (state graph, scheduler, verifier hooks) turned into API surface. Or it can concede the runtime layer to OpenAI and compete on the model plus the coding harness experience alone, which is a defensible position given the 75 percent cache-read cut on Fable 5.1 but a smaller product than the one the Agents API just staked out.

None of the three is free. The one Anthropic has been signaling with Claude Science and the summer harness-is-the-product bet is the first one, and the calendar to catch OpenAI ran out yesterday.

What It Does to Cursor, Devin, and the Coding Harness Layer

Less than the headlines will read, more than the pricing page shows. Cursor, Cognition Devin, Zed, Windsurf, and the long tail of coding harnesses sell three things that the Agents API does not replace: a code-native UI, an opinionated interaction model tuned to the way developers actually work, and a relationship with the customer that includes billing, seat management, and a support surface. The Agents API is a runtime primitive. It is not a product a developer opens in the morning to write code.

What it does is remove the excuse for those companies to build their own orchestration. A team that raised a Series B on the promise of a proprietary agent runtime now has an OpenAI-owned reference implementation with subagents, compaction, and MCP baked in, and any pitch that leans on runtime differentiation has to answer why the buyer should not just consume the Agents API directly. The runtime layer of the pitch deck is commoditized this week. The UI, the model routing, the eval harness, the customer integration work, and the pricing surface are where the differentiation has to move, and the coding harness leaderboard through the rest of the year is going to reflect which teams already had that answer written down.

The AFTA-Adjacent Read

One quiet detail worth naming for anyone on the agent-payments beat: the Agents API does not include a payments primitive. There is no x402 client, no wallet, no ledger, no settlement layer. Cloudflare is on the sandbox partner list, and Cloudflare is also the x402 co-governance partner and the operator of the buyer-side x402 wallet, so the pieces are inside the same partnership map. But the Agents API itself treats money the way every hosted API has treated money for a decade: pay OpenAI, tell OpenAI which sandbox partner to bill through, and settle the rest downstream.

The AFTA-shaped question that is worth asking now is whether the fee floor for a non-OpenAI-native tool call inside an Agents API run trends toward zero-cost coordination plus per-call receipts, or toward the classic API-marketplace shape where the model vendor takes a rake on every third-party call. OpenAI has not signaled either way, and the answer decides whether the agent-payments rail runs above the Agents API or alongside it. We will be watching the fee-floor thesis against the first Agents API pricing update.

Our Take

The interesting choice is the zero-fee decision on the harness. OpenAI could have charged a per-session premium on Agents API runs, the way AWS charges a per-invocation premium on Lambda above the raw EC2 cost, and the market would have paid it. Instead OpenAI put the price at zero and moved the entire margin of the product to the token line and the sandbox line, both of which OpenAI either already owns (tokens) or has offloaded to named partners (sandbox). That is the shape of a company that has decided the runtime is a distribution channel for tokens, not a business unit of its own, and it is a bet that the next five points of API market share are won by making the surface area of the loop as thin as possible on the customer's invoice.

The read for Anthropic is that the harness thesis is now a race with a shipping reference implementation, not a whitepaper thesis with a pending product. The read for third-party harness startups is that the runtime layer of your pitch is done as a moat, and the interaction layer plus the customer relationship is what has to carry the valuation from here. The read for enterprises is that the cost of building a bespoke agent runtime just got harder to justify against a paved road that ships with subagents, MCP, compaction, and a sandbox partner network out of the box.

Practical implication for builders on the API: if you are running an agent workload on any hosted model today, run the numbers on porting one thin slice of it to the Agents API this week. Not to move production traffic, to price the delta. The number you get back is the number that decides how much of your runtime code is a real asset and how much is scaffolding you were paying to maintain until an API vendor built the same thing for free.

Three signposts for the next 60 days. Whether Anthropic ships an equivalent Agents API tier with Prove2Me-shaped orchestration as a first-class primitive, which is the direct test of whether the harness thesis holds as a two-lab race or collapses to an OpenAI product line. Whether the first Agents API pricing update introduces a per-session or per-subagent fee, which is the direct test of whether zero-fee harness is a launch price or a permanent posture. Whether any of the nine named sandbox partners publishes a revenue-share disclosure on Agents API traffic, which is the direct test of where the margin in this stack actually settles. Two of the three fire and the shape of the agent runtime market for 2027 is set on this week's announcement rather than on whatever ships next.