For 48 hours, the release hung in limbo. A Trump administration security review had halted OpenAI’s plans to launch its newest model, and the company’s engineers spent the intervening time making changes, waiting, and preparing for the possibility that the launch would be delayed indefinitely. On July 9, the hold was lifted, and OpenAI shipped both halves of its biggest product release of the year.
The first half is GPT-5.6, now open to the public in three tiers: Sol, Terra and Luna, with reasoning capability increasing at each step. The naming, drawn from the planets, reflects the product structure; the tiers let OpenAI sell the same underlying model at different prices depending on how much thinking a task requires.
The second half is the more consequential bet. ChatGPT Work is an agent product designed to plan, execute and iterate on complex workflows without step-by-step human instruction. Where ChatGPT answered questions, Work takes assignments: it can research a market, draft the analysis, build the presentation and revise it in response to feedback, all with minimal prompting. InfoWorld, reviewing the release, described the shift as the industry moving from “question-answer” to “delegate-deliver,” and the phrase captures the change in how users are expected to interact with the product.
OpenAI said the security review’s concerns centered on model autonomy and content generation, and that the company had resolved the specific issues raised. The company did not detail the changes, but people familiar with the process said the adjustments were limited and did not alter the model’s architecture or its reasoning capabilities.
The same day, OpenAI closed ChatGPT Atlas, the browser it had launched months earlier with the ambition of building an AI-native alternative to Chrome. Atlas’s features are being folded into ChatGPT Work, and the browser itself is being retired. The product move is strategic: instead of spreading its AI capabilities across a browser, a search product and a suite of standalone tools, OpenAI is consolidating everything into a single workspace.
The consolidation reflects a broader argument about where the value lies. A browser competes for habits, and habits are expensive to change; a work platform competes for outcomes, and outcomes are what enterprises pay for. OpenAI’s leadership concluded that its resources were better spent on the latter, and the Atlas shutdown was the visible consequence of that decision.
The timing of the release matters as much as the content. The model ships less than a week after the administration’s review raised questions about autonomous AI, and its clearance is a signal about the direction of policy: the government’s concern was specific, not existential, and once OpenAI addressed it, the product went out. Companies across the industry are reading the episode as a template for how the second Trump administration will treat frontier AI releases.
The tier structure is engineered for the way AI usage actually grows. Sol handles quick queries where speed matters more than depth; Terra adds multi-step reasoning for drafting, analysis and planning; Luna runs the long-horizon tasks, the research projects and the complex builds that an agent product needs to sustain. The boundaries are not fixed, OpenAI said, and the model can be steered toward deeper reasoning when a task requires it. For customers, the design means paying for depth only when depth is needed.
Competition in the agent market is intensifying. Anthropic has pushed its own agentic products into enterprise deployments, and Google has been embedding agents into its search and workspace products. OpenAI’s answer is a single, consolidated platform with a tiered model underneath it, and the company is betting that the combination of a frontier reasoning model and an agent workspace is the product configuration the market will settle on.
The product also forces a conversation about work itself. An agent that can take a task from brief to finished deliverable changes what enterprises buy software for, and it changes what employees are hired to do. Analysts said the most consequential effect may be on the economics of knowledge work: if a tool can draft, revise and present, the marginal cost of producing an analysis falls sharply, and companies will reorganize around that fact.
The enterprise pitch is straightforward. Companies that adopted ChatGPT as a productivity tool are now being offered a system that can run workstreams end to end, and the pricing tiers are designed to grow with usage. Analysts said the release is OpenAI’s most direct attempt yet to convert its consumer ubiquity into enterprise revenue.
For the engineers who spent two years shipping separate products, the July 9 release closed a chapter. Atlas is gone, the model is out, and Work is the platform everything now routes through. The 48 hours of limbo are over, and OpenAI has staked its next quarter on the proposition that users will stop asking questions and start handing over work.
