OpenAI Stopped Selling You a Model. On July 9 It Started Selling You the Finished Job.

Everyone spent yesterday counting benchmark points. GPT-5.6 Sol went public, Luna landed at a dollar of input, and the timeline filled up with Sol versus Grok versus Sonnet leaderboard screenshots. That was the loud announcement. The quiet one, the one that actually tells you where OpenAI thinks the money is going, was the product it shipped alongside the models. It is called ChatGPT Work, and it does not answer questions. It returns finished work.

I think that is the story from July 9, and almost nobody framed it that way. OpenAI did not just refresh its model lineup. It moved the thing it charges you for one full layer up the stack, from the token to the outcome, and it did it in the same week the token tier collapsed toward a dollar. Those two facts are the same fact. Let me walk through it.

What ChatGPT Work Actually Does

ChatGPT Work is an agent that lives inside ChatGPT. You hand it an outcome, not a prompt. It gathers context across your connected apps, breaks the goal into smaller steps, and works through them on its own, staying with a complex project for hours if the job needs it. What it hands back is not a chat transcript. It is the deliverable: spreadsheets, slides, documents, and interactive web apps.

The connective tissue is a new Unified Plugins Directory that puts the third-party integrations in one place. At launch that list runs through Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, Adobe, Zoom, LinkedIn, GitHub, Canva, and Dropbox. Read that list again. It is not a set of data sources for a chatbot to quote. It is the surface area of a knowledge worker's actual desk. OpenAI is not trying to be the tab you ask a question in. It is trying to be the worker you delegate the task to.

It rolled out July 9 for Pro, Enterprise, and Edu subscribers, with Plus and Business plan holders getting access within days. On desktop, OpenAI folded the standalone Codex app into the ChatGPT client on both Mac and Windows, so the coding agent and the work agent now live behind the same window.

The Pricing Is the Whole Point

Here is the detail that made me stop scrolling. ChatGPT Work is not a flat feature of your subscription. OpenAI says plainly that it "is designed for longer, more involved work than a typical chat request, so usage works differently," and that it "follows the same usage structure as Codex." In practice ChatGPT Work draws from a shared agent-consumption pool alongside Codex, ChatGPT for Excel, and Workspace Agents. How much a task burns depends on its size, its complexity, and the model you point at it.

Sit with what that pricing does. It abstracts the token away. You are no longer buying a million tokens of Sol at thirty dollars of output. You are buying a slice of a pool that gets spent when an agent goes and finishes something for you. The unit of purchase is no longer the token. It is the job. And once the unit is the job, you stop opening a spreadsheet to compare per-million rates across labs, because the number you care about is how many finished deliverables the pool buys you this month.

That is not an accounting quirk. That is a moat under construction. Per-token pricing is brutally legible: any buyer can line up Sol, Grok, and Sonnet in a row and pick the cheap one. Consumption-pool pricing is deliberately illegible, and OpenAI reached for it at the exact moment legibility started working against the incumbent.

Why Now: The Token Just Became a Commodity

Look at what the token tier did in the 48 hours around this launch. Grok 4.5 shipped July 8 at two dollars input and six dollars output. Sonnet 5 is running introductory pricing at two and ten. Then on July 9 OpenAI itself released Luna, the small tier of the GPT-5.6 family, at one dollar input and six dollars output. Three capable agent-grade models, all from different labs, all clustered inside a couple of dollars of each other at the bottom.

When your cheapest capable model and two rivals' models all sit within a few dollars of each other, the token stops being where you differentiate. It becomes salt. Necessary, cheap, and interchangeable. You can watch that convergence yourself on our models tracker and benchmarks page. The direction of travel has been obvious for a year. What changed this week is that the leader stopped pretending the token was still the product.

So OpenAI did the only rational thing a company with the strongest distribution in the category can do. It kept the token race going, priced Luna right into the mud with Grok, and then quietly moved its own margin story up to a layer where nobody can run a clean price comparison. You cannot benchmark "a finished quarterly deck" the way you can benchmark a thousand tokens of output.

Codex Was the Dress Rehearsal

None of this came out of nowhere. Codex was the test case, and the numbers explain the confidence. Weekly Codex usage passed five million people, a jump of roughly 400 percent across 2026. That is not a demo. That is a habit forming at scale, and it is a habit built entirely around delegating a whole task, not sampling a token.

The other tell was the acquisition. In June, OpenAI moved to buy Ona, a German startup whose entire product is a secure, persistent cloud where an agent keeps working after the developer who kicked off the task has closed the laptop and gone home. Ona says productive use of its agents among enterprise clients climbed 13-fold in 2026. You do not buy a company like that to make chat faster. You buy it because your product is now a worker that needs somewhere to keep working when the human steps away. ChatGPT Work staying with a project "for hours" is the consumer-facing face of exactly that capability.

Where This Fits the Pattern We Have Been Tracking

Regular readers know the thesis I keep coming back to: the model is not the product, the workflow is. We watched AWS and Microsoft stand up billion-dollar consulting arms and lift the forward-deployed-engineer playbook, selling embedded pods that deliver outcomes rather than API access. We watched Microsoft start routing its own Excel and Outlook prompts onto in-house models to squeeze the supplier. Same move, different altitude.

ChatGPT Work is that same move aimed at the individual seat instead of the enterprise contract. The hyperscalers are selling outcomes by sending you engineers. OpenAI is selling outcomes by sending you an agent. Both are refusing to let the buyer's attention rest on the model layer, because the model layer is where the price war lives and where nobody keeps a durable lead for more than about 24 hours. That is a claim you can now verify with a stopwatch. Grok 4.5's price advantage lasted exactly one morning.

The Catch

I am not selling this as a clean win. Moving up the stack means OpenAI is now competing with the very apps in its own plugin directory. If ChatGPT Work returns a finished spreadsheet, that is a shot at every tool that used to own the spreadsheet. Salesforce, Adobe, and Canva are on that launch list as integrations today. Ask yourself how long a company happily feeds context to an agent that is learning to produce the exact artifact it sells. The plugin directory is a truce, and truces in this market have a short shelf life.

The consumption-pool pricing cuts both ways too. Illegible pricing protects margin, but it also makes buyers nervous, because a pool that drains at a rate you cannot predict is a budget you cannot forecast. Codex users have already felt this. The first enterprise that gets a surprise overage on a month of long-running agents will make sure everyone hears about it. If you want to model your own exposure before that happens, our cost calculator is the place to start.

Our Take

Watch the byline war less and watch the billing unit more. The single most important thing OpenAI shipped on July 9 was not a benchmark point. It was the decision to price the job instead of the token, timed to the exact week the token stopped being worth pricing. That is a company reading its own commodity curve correctly and refusing to die on it.

Three things I am watching over the next 90 days. First, whether Anthropic and Google answer with their own outcome-priced agent products or hold the line on per-token API purity. Second, whether the consumption pool produces a public billing-shock story big enough to slow enterprise adoption. Third, whether the apps in that plugin directory stay friendly once ChatGPT Work starts producing the deliverables they were built to sell.

The token got cheap this week. That was always going to happen. The interesting question was what the leader would do the day it finally did, and now we know. It moved the price tag off the token and onto the outcome, and it did it while everyone was still arguing about leaderboards.