OpenAI Launches GPT-5.6, ChatGPT Work, and a Unified Desktop App: The Next Chapter for ChatGPT

Today's event (Thursday, July 9, 2026) was not a model launch, but "the next chapter for ChatGPT" as OpenAI described it in its livestream. After a limited preview subject to a government review, the GPT-5.6 family was released to the public, but it came loaded with a suite of tools reshaping the entire product: a new "ChatGPT Work" agent, a unified desktop app integrating Codex, and a hosted sites service. Let's detail what launched, its features, and what the tests say.

The Three Models: Sol, Terra, and Luna

GPT-5.6 breaks the numeric naming tradition with three models whose names denote "durable capability tiers": Sol the flagship for frontier reasoning and long-horizon agentic work (with a stronger Sol Ultra, and a Sol Pro for the highest quality), Terra the balanced model OpenAI describes as competitive with GPT-5.5 at half the cost, and Luna the fastest and cheapest. The release introduces two notable features: "max reasoning effort," giving Sol the longest time to think, and "ultra mode," which distributes work across subagents to accelerate complex tasks. OpenAI announced that GPT-5.4 will be retired on July 23, while GPT-5.5 remains available.

The Standout Tool: The ChatGPT Work Agent

The most important thing accompanying the model is "ChatGPT Work," an agent powered by GPT-5.6 designed to automate office work. It can work "continuously for hours" on complex projects, and translates broad goals a user describes into completed work. In practice, it gathers context from connected apps, files, and workflows, then produces finished materials: documents, spreadsheets, presentations, and even web apps. It works across web, phone, and computer, starting with the Mac and Windows apps for all user tiers, with web to follow. For those following along, the direct resemblance to Anthropic's Claude Cowork tool, whose mobile expansion we covered two days ago, is clear — the competition over "office work automation" is intensifying.

A Unified Desktop App... and Codex Integration

OpenAI launched an entirely new desktop app that unifies three capabilities in one interface: Chat, Work, and coding (Codex). This achieves the promised "merging of Codex into ChatGPT": a single app where the user can set preferences to favor Codex functionality, including the app icon. Thus the desktop experience shifts from multiple windows to a single work hub. OpenAI also launched a "Hosted Sites" service for all paid users, to publish what is built directly.

The Numbers and Tests: What OpenAI Announces

OpenAI presents strong numbers, which must be read as coming from the model's maker rather than a neutral party. Per OpenAI, Sol posts record performance on Terminal-Bench 2.1 (about 88.8%, and Sol Ultra about 91.9%), surpassing what it attributes to the Claude Mythos 5 model (88.0%). CEO Sam Altman told CNBC that Sol is "54% more token efficient" on agentic coding tasks — an important figure in an era when every enterprise thinks about spend versus value. OpenAI also claims improvement in biology (SecureBio) and cybersecurity, and a leap in "design sense": the ability to inspect and visually refine the rendered interface, not just generate its code. None of these numbers have been independently verified yet.

Testers' Testimonials: A Comparison With Fable

The model drew enthusiastic reviews from early testers, with notable comparisons to Anthropic's Fable model. T3 Chat platform head Theo Browne wrote that "Sol is world-leading in computer use." But the picture is not decisive in favor of one side: some testers see Fable as having "greater raw intelligence," while GPT-5.6 is viewed as a "more reliable model for regular tasks." Every CEO Dan Shipper put it eloquently: "If you need to get across the galaxy, use Fable." This balance matters: each model has a strength, and the choice depends on the task rather than on "the absolute best."

The Government Dimension: A Recurring Precedent

The deeper story is that GPT-5.6 was delayed behind a government gate. OpenAI previewed the model to the US government, and at its request started with about twenty approved partners before expanding, because of Sol's cyber capabilities. This exactly parallels what happened with Anthropic's Fable 5 and Mythos 5. The significance, as Axios summed up, is that Washington treats the most powerful models "as products that need government review." OpenAI was clear in its reservation: "We don't believe this process should become the long-term default." It says it believes Sol is "better at helping people find and fix vulnerabilities than carrying out full attacks," and that its capabilities do not reach the "critical" level.

All of this comes within a notably accelerating cadence: OpenAI has compressed its release cycle to under sixty days. This shifts how teams deal with models, from "committing to one for a year" to "managing rolling upgrades." And the practical lesson echoing the rise of alternative models: the value is in building an "abstraction layer" that lets you swap in the winning model within a week, not in betting on a single model. The battle today is no longer over "the most powerful model" alone, but over who offers the most integrated tool suite around that model — and here OpenAI shifted from selling "intelligence" to selling a complete "work platform."