OpenAI ChatGPT Work: The Super App Merging Codex

OpenAI has folded its two flagship consumer products into one. On July 9, 2026, the company merged ChatGPT and its coding tool Codex into a single desktop application and introduced ChatGPT Work, an agentic workspace aimed squarely at people who are not developers. The release lands the same week OpenAI completed the general rollout of GPT-5.6, the three-tier model family that now powers the combined app — and it escalates a product war with Anthropic that has moved, in a matter of days, from coding tools to everyday knowledge work.

The framing is a deliberate one. For most of the past two years, the frontier labs sold two separate things: a chatbot for the general public and a coding agent for engineers. OpenAI’s July 9 move erases that line. The same agent that writes and reviews software can now build a slide deck, draft a spreadsheet, or assemble a website — and it lives in one place.

What actually shipped

The launch bundles several changes that had been rolling out piecemeal:

• Codex is now inside ChatGPT. The standalone Codex app has been merged into the ChatGPT desktop client, so a single window handles both conversation and code. OpenAI also moved inline diff editing and pull-request review directly into Codex, tightening the loop between an agent proposing a change and a developer accepting it.
• ChatGPT Work is the headline product. Powered by GPT-5.6, it is pitched at non-coders who want the capabilities of an AI coding agent applied to office work — generating documents, presentations, and websites from natural-language instructions rather than code.
• Shared usage and billing. ChatGPT Work and Codex now draw from the same pool. Work usage inside ChatGPT is metered on the same credits, rate limits, and pricing as Codex, and both are included across the Free, Go, Plus, Pro, Business, Edu, and Enterprise plans.

The model layer underneath is the GPT-5.6 family that OpenAI began previewing earlier in the month. It splits into three tiers: Sol, the flagship aimed at the hardest agentic and reasoning tasks; Terra, a mid-tier model OpenAI has positioned at roughly GPT-5.5 performance for about half the cost; and Luna, the fastest and cheapest option. ChatGPT Work routes across the three depending on the job, which is how OpenAI can offer long-running agentic sessions without the token bill becoming prohibitive.

The pricing mechanics

OpenAI has been steadily shifting Codex — and now the merged app — toward API-style token pricing. Under the current rate card, customers still buy credits, but consumption is measured against input, cached input, and output tokens rather than a flat per-seat allotment. The Business plan runs $25 per user per month billed annually, and teams on Business and Enterprise can now add Codex-only seats on a pay-as-you-go basis, with no fixed seat fee and no rate limit — usage is billed purely on tokens consumed.

For enterprises, the controls are the part that makes the app deployable at scale: SCIM provisioning, enterprise key management (EKM), role-based access control, domain verification, audit logs, and usage monitoring through a Compliance API. Those features are table stakes for selling into regulated industries, and their presence signals that ChatGPT Work is meant to sit alongside — or replace — existing productivity suites inside large organizations, not just live as a consumer novelty.

Anthropic’s counter

OpenAI is not moving into an empty field. Just two days earlier, on July 7, Anthropic launched Claude Cowork on mobile and the web, starting with its Max subscribers at $100 per month. Cowork’s pitch is nearly identical to ChatGPT Work’s: hand an agent a knowledge-work task — spanning documents, spreadsheets, and presentations — and let it run. Its headline feature is that agentic sessions can now continue in the background even when a user’s laptop is closed or offline, removing the long-standing constraint that tied a running agent to an always-on desktop.

The competitive symmetry is striking. Both companies spent the first half of 2026 racing on raw model capability — Anthropic with its Claude line, OpenAI with the GPT-5 series — and both have now pivoted the same week to packaging that capability as an everyday work assistant for non-engineers. Anthropic underscored the moment by extending free access to its Fable 5 model through July 13, keeping its most capable model in front of users precisely as OpenAI tried to dominate the news cycle.

The category itself is not new — Microsoft has pushed Copilot into its Office applications for years, and a wave of startups has built agentic harnesses on top of both labs’ APIs. What changed on July 9 is that the model makers are now selling the finished application directly, competing with the ecosystem of tools built on their own models. That tension has been building across the AI coding-assistant market, where OpenAI and Anthropic increasingly ship first-party products that overlap with their customers’ businesses.

Why merge the apps at all

The strategic logic is about surface area and retention. ChatGPT’s consumer reach is enormous, but its per-user economics are thin; Codex’s engineering users are far fewer but spend heavily on tokens for long agentic runs. Merging them lets OpenAI push its highest-value capability — autonomous, tool-using agents — in front of its largest audience, and it collapses two engineering roadmaps into one.

It also reframes what “using ChatGPT” means. A chatbot answers a question and ends the turn. An agent takes a goal, plans, calls tools, edits files, and reports back — the same shift that has defined the broader move toward autonomous AI agents and the harnesses being built to orchestrate them, from OpenAI’s Codex to third-party frameworks like Databricks’ Omnigent. By putting that machinery inside the app hundreds of millions of people already open daily, OpenAI is betting it can convert casual chatbot users into agent users, and agent users into paying seats.

There is a defensive dimension too. As models commoditize — Terra reportedly matching last generation’s flagship at half the cost — the durable moat shifts from the weights to the product: the workflow, the integrations, the enterprise controls, and the habit. Owning the application layer is how a lab keeps a customer even when a rival’s model briefly tops a benchmark, a dynamic on vivid display in the same week that saw fresh model launches trade the top spots on independent leaderboards.

The open questions

Several things remain unclear. OpenAI has not detailed how ChatGPT Work handles the reliability problems that plague agentic systems — hallucinated edits, silent failures on long tasks, and the difficulty of verifying an agent’s output when the user is, by design, not an engineer. Pushing autonomous agents to non-technical users raises the stakes on exactly the failure modes that expert users can catch and correct.

Pricing is also in flux. The move to token-based billing gives enterprises a clearer view of spend, but it makes costs less predictable for individuals running long agentic sessions — the opposite of the flat-rate simplicity that made ChatGPT’s consumer subscription easy to reason about. How OpenAI balances that against Anthropic’s seat-based Max tier will shape which app teams standardize on.

What it means

The merger marks the moment the AI product war moved decisively past the chatbot. OpenAI and Anthropic are no longer competing only to have the smartest model; they are competing to own the application where work actually happens — and both concluded, within the same week, that the prize is the non-coder who has never opened a terminal.

Who wins. OpenAI, if it converts scale into paid seats. Bundling Codex’s agentic capability into the app that already has the largest consumer footprint is the single most efficient distribution move available to it, and the enterprise controls make the app sellable to the buyers who spend the most. Anthropic is well-positioned to hold the high end, where Claude’s coding reputation and Cowork’s background-execution feature give it a credible, differentiated pitch.

Who should be nervous. The startups that built businesses as a thin layer over these labs’ APIs. When the model maker ships the finished product — one app for chat, code, documents, and websites — the value of a wrapper shrinks. Microsoft, too, now faces first-party competition for the office-productivity workflow that Copilot was built to own.

What to watch next. Three things. First, adoption metrics: whether non-coders actually use ChatGPT Work for real deliverables or treat it as a novelty. Second, reliability in the wild — the first high-profile failure of an agent editing a non-engineer’s work will test how much trust the category can bear. Third, Anthropic’s response after July 13, when Fable 5’s free window closes and both labs settle into the pricing that will define the everyday-work AI market for the rest of the year. The chatbot era gave way to the agent era this week; the question now is whose agent people open in the morning.