On September 10, 2026, OpenAI made two announcements that look unrelated until you put them side by side. It opened the Agents API to all developers in public beta, exposing the same managed harness that runs its Codex coding agent and ChatGPT for Work. And it launched ChatGPT for Financial Services, a product aimed squarely at the daily grind of investment banking and equity research. One move goes down the stack toward infrastructure other people build on. The other goes up the value chain into a specific, lucrative profession. The common thread is a company trying to own more than the model.
OpenAI
Microsoft
NVIDIAWhat the Agents API actually replaces
The interesting thing about the Agents API is not that OpenAI built an agent. It is that OpenAI is selling the plumbing every serious agent team has been rebuilding by hand. A useful agent needs a harness: something that manages a long context window without losing the thread, calls tools reliably, spawns and coordinates subagents, and keeps a task alive for hours or days without falling over. That harness, not the raw model, is where most of the engineering effort in agent products has actually gone.
OpenAI is now offering it behind a single API call. The service handles context compaction, tool orchestration and recovery, and it lets an agent break a job into pieces and hand them to subagents that run in parallel, each with its own context, while a main agent coordinates and merges the results. Developers choose where the code runs: an OpenAI-hosted sandbox, their own infrastructure, or one of nine named partners including Cloudflare, E2B, Modal, Oracle and Vercel. There is no separate fee for the API itself; you pay for model tokens, the tools invoked, and sandbox minutes.
The model was the product in 2023. In 2026 the harness is the product, and the model is a component inside it.
On what the Agents API signals
That pricing choice is a strategy in disguise. By charging only for consumption and giving the orchestration away, OpenAI makes it cheaper to start an agent on its stack than to maintain a homegrown harness. The subtle cost is gravity: the more of your agent's lifecycle runs inside one vendor's managed environment, the harder it becomes to move it. Convenience and lock-in are, as usual, two readings of the same feature.
The finance product is a template, not a one-off
ChatGPT for Financial Services runs on GPT-6 Astra and ships with data built in: earnings transcripts, financial statements, company fundamentals and private-company records sourced from providers such as Daloopa, PitchBook and LSEG News, with integrations spanning S&P Capital IQ, MSCI, Moody's, Dow Jones Factiva and FactSet, plus more than fifty MCP connectors into tools like Datasite, Box and Preqin. It was shaped with Morgan Stanley and Evercore as design partners, and it is pointed at the specific tasks that fill a junior banker's week: normalizing a P&L, building a valuation model, screening potential buyers, and turning all of it into client-ready material.
The model behind the finance product, on OpenAI's own test
OfficeQA Pro score, higher is better. ChatGPT for Financial Services runs on GPT-6 Astra.
OfficeQA Pro is an OpenAI benchmark; the result is the vendor's own reported figure, not an independent evaluation.
OpenAI reports that GPT-6 Astra scores 69.9% on its internal OfficeQA Pro benchmark against 60.2% for the previous GPT-5.6 Sol. That is a meaningful gain on the vendor's own test, and it is worth reading with the usual caution: a benchmark authored by the company selling the model tells you the direction of progress, not how the system performs on a specific desk's messy inputs. The honest way to evaluate a product like this is on a bank's own historical deals, not on a leaderboard.
The reason this matters beyond finance is that it is a repeatable pattern. Take a frontier model, wrap it in a vertical's proprietary data and connectors, tune it to that vertical's exact tasks, and sell it to institutions that will pay far more for a finished workflow than for API access. Law, healthcare, accounting and consulting all have the same shape: high-value knowledge work, structured data behind paywalls, and expensive junior labor doing repetitive analysis. Finance is simply the first vertical where the data partnerships and the willingness to pay lined up.
Two moves, one thesis
Put together, the day tells a coherent story. The Agents API is a bet that OpenAI can be the default place developers build agents, capturing consumption from thousands of applications it will never have to design. ChatGPT for Financial Services is a bet that OpenAI can also capture the high end directly, selling finished products into professions where the margins are richest. The platform play widens the base; the vertical play deepens the take. A company confident in its model would do one. A company trying to become an operating system for work does both at once.
For teams deciding how to build on this, the strategic question is less which vendor is ahead this quarter and more how much of your workflow you are willing to fuse to a single provider's harness and hosted environment. The counter-discipline is to keep the orchestration layer yours: to route work across models, to run agents in parallel without surrendering the whole lifecycle to one stack, and to keep your data and outputs portable. That model-agnostic posture is the design principle behind platforms like Metir, which coordinates multiple models and agents while leaving the user in control of where the work runs and where it can move next. The Agents API makes starting on one stack almost free. The work of staying free to leave is the part that pays off later.
Neither announcement settles who wins the agent era. What they show is a clear read of where the value is migrating: away from the raw model, toward the harness that makes agents reliable and the vertical products that make them worth paying for. On September 10, OpenAI planted a flag on both.
Sources:
- Introducing the Agents API (OpenAI)
- OpenAI Launches the Agents API in Public Beta (MarkTechPost, Sept 10, 2026)
- OpenAI Launches ChatGPT for Financial Services With Built-In Data (Unite.AI)
- OpenAI targets work of Wall Street junior bankers with new ChatGPT for Financial Services (CNBC, Sept 10, 2026)
- OpenAI Debuts ChatGPT for Financial Services (Bloomberg, Sept 10, 2026)
Image credits
- Hero: New York Stock Exchange facade, May 2024. Photo by Ajay Suresh, Wikimedia Commons, licensed CC BY-SA 4.0.
- In-body: New York Stock Exchange and Wall Street, August 2017. Photo by King of Hearts, Wikimedia Commons, licensed CC BY-SA 4.0.