GPT-5.6 and the New ChatGPT Desktop App: Complete Guide

OpenAI launched three new models on July 9. The company also replaced its desktop strategy.

The GPT-5.6 release introduced Sol, Terra, and Luna. Each model targets a different combination of intelligence, speed, and price. Alongside them, OpenAI launched ChatGPT Work, merged Codex into ChatGPT, added a built-in browser, expanded computer use, introduced shareable Sites, and gave scheduled agents more access to your actual workflow.

The model upgrades matter. The larger bet is that ChatGPT can become the place where work gets planned, delegated, reviewed, and delivered.

Think of GPT-5.6 as the engine. The new ChatGPT desktop app is the vehicle OpenAI hopes you will spend your day driving.

That vehicle now has a striking number of controls. Chat, Work, Codex, Sites, Scheduled Tasks, plugins, reasoning levels, Pro mode, Max, and Ultra all live in or around one product.

Somewhere, a beginner opened the model picker and quietly returned to Microsoft Word.

This guide explains what each part does, what changed from the old ChatGPT and Codex apps, and how to choose the right model without conducting a benchmark study before lunch.

What OpenAI launched

The announcement had four major parts:

GPT-5.6, a family of three models called Sol, Terra, and Luna.
ChatGPT Work, an agent designed for longer projects across files, apps, and websites.
A unified desktop app, combining ChatGPT and Codex on Windows and Mac.
New work surfaces, including Sites, Scheduled Tasks, a built-in browser, computer use, and a unified plugin directory.

OpenAI positioned the release around a simple progression. Chat answers questions. Work handles broader projects. Codex exposes more technical controls for software and computer-based tasks.

The company says Work can stay with a project for hours, gather context from connected apps, split the job into smaller tasks, and return finished documents, spreadsheets, presentations, analyses, or websites. (openai.com)

OpenAI’s product page describes ChatGPT Work as a persistent agent powered by GPT-5.6. It can start on desktop, continue running while you are away, and let you check its progress from mobile. (openai.com)

The result resembles a workbench more than a chatbot. The prompt remains the starting point, but the expected output is increasingly a completed project.

GPT-5.6 is a model family, not one model

GPT-5.6 introduces durable capability tiers instead of one model followed by a trail of Mini, Turbo, Pro, and Preview labels.

GPT-5.6 Sol

Sol is the flagship. It is intended for ambiguous problems, long-running projects, advanced coding, research, strategic analysis, and tasks where judgment matters more than speed.

OpenAI says Sol set new highs on several agentic, coding, browsing, cybersecurity, and scientific evaluations. On Agents’ Last Exam, which tests long-running professional workflows, Sol scored 53.6. Claude Fable 5 scored 40.5 in OpenAI’s comparison.

OpenAI also reported that Sol Max finished Artificial Analysis Intelligence Index tasks 61% faster than Fable at roughly half the estimated cost. These are vendor-published comparisons, so they should be treated as evidence rather than a final verdict. (openai.com)

Use Sol for work such as:

Reviewing a complicated codebase and implementing a repair.
Building a research report from dozens of sources.
Creating and refining an application across several iterations.
Analyzing a strategy problem with incomplete or contradictory evidence.
Coordinating several subagents across one larger project.

GPT-5.6 Terra

Terra is the balanced daily model. OpenAI positions it below Sol in raw capability, but closer to the price and speed needed for regular professional work.

Terra makes sense when the task has several steps but the outcome is clear. It can draft, analyze files, build straightforward software, use tools, and complete many agentic workflows without paying Sol prices.

For most people exploring Work for the first time, Terra is the sensible starting point.

Use Terra for:

Weekly reports.
Document and spreadsheet analysis.
Routine software changes.
Research with clear boundaries.
Content production and operational workflows.
Repeated automations that still require some judgment.

GPT-5.6 Luna

Luna is the high-volume model. It prioritizes speed and cost efficiency.

Luna is suited to clearly defined tasks that may run hundreds or thousands of times. OpenAI says it nearly matches GPT-5.5’s peak performance on some professional evaluations while costing less.

Use Luna for:

Classification and extraction.
High-volume customer-support workflows.
Processing standardized forms or documents.
Basic research collection.
Repetitive agent steps.
First-pass analysis before escalation to a stronger model.

OpenAI’s developer guidance summarizes the lineup neatly: Sol for frontier capability, Terra for a balance of intelligence and cost, and Luna for efficient, high-volume work. (developers.openai.com)

How much the GPT-5.6 models cost

For developers using the API, OpenAI priced the family per one million tokens:

Sol: $5 input and $30 output.
Terra: $2.50 input and $15 output.
Luna: $1 input and $6 output.

The sticker price only tells part of the story. Models that finish a workflow using fewer tokens, fewer tool calls, and fewer repair attempts may cost less overall even when their token rate is higher.

OpenAI built much of its GPT-5.6 pitch around this idea: the useful metric is cost per successful task, not cost per token.

That distinction grows important when an agent works for several hours, calls dozens of tools, spawns subagents, and repeatedly checks its own work. A small efficiency difference at each step can become a large bill by the end. (openai.com)

Max, Ultra, and Pro are three different things

OpenAI now gives users several ways to increase the amount of work a model performs. The names overlap enough to make the interface feel like the difficulty screen from a video game.

They control different layers.

Max reasoning

Max gives one model additional time to reason, test alternatives, and verify its answer.

GPT-5.6 supports reasoning levels including none, low, medium, high, xhigh, and max. OpenAI recommends medium as a balanced starting point and reserves max for difficult, quality-first tasks.

Max works well when one agent can solve the problem, but that agent needs to explore more possibilities or check its work carefully.

Examples include:

A difficult financial model.
A complex code review.
A strategic recommendation with several constraints.
A mathematical or scientific problem with a checkable answer.

OpenAI recommends testing the same reasoning level you used with an older model and then testing one level lower. GPT-5.6 may achieve similar quality with fewer reasoning tokens. (developers.openai.com)

Ultra mode

Ultra turns one project into a team project.

OpenAI says Ultra coordinates four agents by default. The agents can handle independent workstreams in parallel before one system combines the results.

A market-research task could assign separate agents to customers, competitors, pricing, and regulations. A software project could divide architecture, implementation, testing, and review.

Ultra works best when the task separates cleanly. Four agents researching the same narrow question may create more duplication than insight.

It can improve quality and reduce the time you wait, but it also consumes more tokens. Ultra is available in ChatGPT Work for Pro and Enterprise users and in Codex for Plus plans and above. (openai.com)

Pro mode

Pro applies more model work before returning one final response.

In the API, Pro is an execution mode rather than a separate model. Developers choose Sol, Terra, or Luna, then enable Pro independently from the reasoning level.

OpenAI recommends Pro when a marginal increase in reliability has meaningful value. That could include high-value code review, optimization, deep analysis, or decisions with measurable success criteria.

Routine work should remain in standard mode unless testing proves Pro improves the result enough to justify its extra latency and cost. (developers.openai.com)

The simplest mental model:

Model tier determines the base capability.
Reasoning level controls how much one model thinks.
Ultra coordinates several agents.
Pro spends more model work producing one final answer.

The new ChatGPT desktop app is built on Codex

OpenAI’s desktop change sounds simple until you try explaining which application became which.

The standalone Codex app now becomes the primary ChatGPT desktop app. Existing Codex users can update normally and keep their projects, settings, and workflows. They can also make Codex the default opening view and, on macOS, keep the Codex icon.

The older ChatGPT desktop application is being renamed ChatGPT Classic.

The new app contains three central work modes:

Chat for questions, conversation, and quick tasks.
Work for broader projects across files, apps, and websites.
Codex for technical workflows with more detailed controls.

All three are available in the desktop app on Free, Go, Plus, Pro, Business, Enterprise, and Edu plans. The updated app is available globally for Windows and Mac. Work on web and mobile began rolling out first to Pro, Enterprise, and Edu accounts, followed by Plus and Business. (openai.com)

9to5Mac’s launch coverage described the transition as Codex becoming the new ChatGPT app while retaining a Codex-focused mode for people who want the technical detail.

That product choice reveals OpenAI’s direction. The company built a powerful agent for developers, watched people use it for writing, research, planning, and operations, then placed the same underlying system behind a broader name.

Codex remains inside. ChatGPT becomes the front door.

What ChatGPT Work does

ChatGPT Work is the most important new surface for non-developers.

It is designed for tasks with several stages, several sources, or a longer runtime than a normal chat. You give it a goal, let it collect context, review its plan, and then watch or redirect the work.

A typical workflow looks like this:

Describe the outcome. Tell Work what should exist when the job is finished.
Connect the context. Give it access to relevant files, folders, emails, calendars, chats, and business tools.
Review the plan. In Plan mode, it can ask questions and propose a sequence before acting.
Let it execute. The agent can research, edit files, use websites, call plugins, and create deliverables.
Check progress. You can review intermediate work, change direction, or answer questions.
Approve important actions. External writes, destructive changes, and other consequential steps can require your confirmation.
Receive the finished output. That may be a document, presentation, spreadsheet, analysis, application, or updated workflow.

OpenAI says ChatGPT Work can combine material from Slack, Teams, Gmail, Outlook, SharePoint, Google Drive, Microsoft 365, Salesforce, project trackers, and other connected systems.

The company says the plugin directory contains more than 1,400 integrations. Users can let the agent choose a relevant plugin or call one directly by typing “@” followed by the app name. (openai.com)

The practical difference from normal chat is persistence. A chat waits for your next message. Work can continue gathering, building, and revising while you handle something else.

Plan mode is the safest place to begin

Plan mode lets ChatGPT gather context, ask clarifying questions, and show its proposed steps before execution.

This is especially useful when:

The request involves several systems.
You have not used the workflow before.
The agent may edit important files.
The outcome has several valid interpretations.
You want to divide work between the agent and a human.

A good Work prompt defines four things:

Outcome: What should exist at the end?
Sources: Which files, apps, and websites can it use?
Boundaries: What may it change without asking?
Success criteria: How will you judge whether the work is complete?

OpenAI’s GPT-5.6 prompting guidance says the model has become more proactive and persistent. It recommends clear approval boundaries instead of repeating “ask first” throughout the prompt.

A compact policy can allow reading files, inspecting logs, editing in-scope material, and running non-destructive tests. Purchases, external messages, deletions, and major scope changes can remain gated behind approval. (developers.openai.com)

Computer Use can finally work in the background

Previous computer-use demos often required the user to sit and watch an agent move through an interface one click at a time.

The new desktop app lets ChatGPT operate across apps, files, and the browser in the background. OpenAI says it can click, type, move files, and complete steps while the user continues working elsewhere.

Computer Use can run as a one-time action or as part of a Scheduled Task. (openai.com)

That turns several previously awkward workflows into realistic candidates for automation:

Downloading a report and adding its numbers to a spreadsheet.
Moving attachments into the correct project folders.
Copying approved content into a publishing tool.
Checking dashboards and updating a weekly presentation.
Collecting information from systems without a direct plugin.
Running a recurring workflow across several desktop applications.

Background execution improves leverage, but it also increases the cost of a mistake. A slow agent you are watching may be annoying. A fast agent making the wrong change in three systems can become a much larger problem.

Start with reversible tasks. Limit the folders and accounts it can access. Require approval before sending, deleting, buying, publishing, or changing permissions.

The built-in browser replaces part of Atlas

The new ChatGPT app includes a browser with multiple tabs and tighter integration with Work and Codex.

It can gather information, open online documents, interact with web applications, and keep browser activity inside the same environment as the rest of the project.

OpenAI is also updating its Chrome extension to place ChatGPT in Chrome’s sidebar. That offers a lighter option for people who want ChatGPT beside their regular browsing without switching browsers.

Those changes come with a casualty. OpenAI has begun sunsetting ChatGPT Atlas, its standalone agentic browser.

9to5Mac reported that the targeted Atlas deprecation date is August 9, with transition details expected through the application and email. OpenAI says the new browser and extension incorporate lessons from Atlas users.

The product logic is clear. OpenAI no longer needs to persuade people to adopt an entirely new browser before they can use agentic browsing. The browser can now sit inside ChatGPT, while the Chrome extension reaches users who prefer their existing setup.

Sites lets users turn work into an interactive website or web application and publish it through a URL.

OpenAI suggests use cases including:

Live dashboards.
Project trackers.
Launch calendars.
Internal portals.
Interactive reports.
Working prototypes.
Small business tools.
Data visualizations.

The site can draw from connected sources and update when its underlying information changes. Users can test it inside ChatGPT before sharing it privately or publicly. (openai.com)

This removes one of the least glamorous barriers in AI-assisted software creation.

Building a prototype has become easy enough for non-developers. Deploying it still often requires GitHub, hosting accounts, environment variables, domains, and a string of decisions that have little to do with the original idea.

Sites creates a direct path from “make this” to “send me the link.”

That makes ChatGPT more competitive with products such as Replit, Lovable, Bolt, and other prompt-to-app platforms. It also expands the addressable audience for Codex-style development beyond people who identify as programmers.

Scheduled Tasks turn projects into recurring jobs

Scheduled Tasks can run once, repeat on a schedule, trigger when an event occurs, or monitor for a change.

OpenAI’s examples include:

Reviewing Slack updates and refreshing a meeting agenda.
Checking websites and dashboards each morning.
Monitoring customer feedback and prioritizing recurring themes.
Updating a presentation when new feedback arrives by email.
Preparing materials before recurring customer meetings.

Schedules can use connected apps, the browser, and Computer Use. Users can check status or review outputs from their phones. (openai.com)

This is the bridge from an impressive conversation to a dependable workflow.

A one-time research report saves an afternoon. A task that refreshes the report every Monday can change how the team operates.

The best first automation has three qualities:

The input arrives predictably.
The output follows a stable format.
Mistakes are visible and reversible.

Weekly summaries, recurring research, meeting preparation, and status reports fit that pattern. Sending contracts, changing account permissions, or publishing unreviewed statements does not.

Codex received its own upgrades

Codex remains available inside the desktop app for developers and technical users.

Its launch-day additions include:

Direct editing of Markdown and code.
Inline annotations and revision requests.
GitHub pull-request review in the side panel.
Support for multiple repositories inside one project.
Faster Computer Use powered by GPT-5.6.
Clearer activity and progress displays.
Shared plugins with ChatGPT Work.
Continued access to desktop projects through mobile.

The distinction between Work and Codex appears to be one of abstraction.

Work hides more of the technical machinery and emphasizes goals, plans, and finished business outputs. Codex exposes repository details, diffs, terminals, environments, and implementation activity.

People who build software may prefer Codex as their default view. People who need research, presentations, analysis, and app automation may spend more time in Work.

Many users will move between both during one project.

GPT-5.6 changes how agents use tools

Several of GPT-5.6’s most consequential upgrades live beneath the interface.

Programmatic Tool Calling

Traditional agents often call one tool, send the result back to the model, decide what to do next, and repeat.

Programmatic Tool Calling lets GPT-5.6 write lightweight JavaScript that coordinates eligible tools inside a hosted runtime. It can process intermediate results, filter unnecessary data, and pass only the useful material forward.

This reduces model round trips and can lower token use for bounded, tool-heavy workflows.

Imagine collecting thousands of customer records. The agent can use a program to filter them before asking the model to make a judgment. The expensive reasoning model sees the relevant 50 records rather than every raw entry. (openai.com)

Multi-agent workflows

Developers can access a beta multi-agent capability through the Responses API.

One GPT-5.6 instance coordinates several subagents, assigns independent workstreams, and synthesizes their findings. This is the API counterpart to Ultra mode.

Parallelism helps when work divides naturally. It has less value when every step depends on the result of the previous one.

Persisted reasoning

GPT-5.6 can reuse reasoning items across turns.

For long projects with stable goals and assumptions, developers can preserve relevant reasoning rather than rebuilding the same analysis after every message. That can improve continuity and caching efficiency.

Explicit prompt caching

Developers can mark reusable prompt prefixes for caching instead of relying only on automatic detection.

Cache writes cost 1.25 times the normal input rate, while cache reads receive a 90% discount. This favors applications that reuse the same large instructions or background context many times.

The system benefits recurring agents more than one-off chats. A workflow that repeatedly loads the same company policies, schemas, or tool instructions can amortize that context across many runs. (openai.com)

GPT-5.6 may prefer shorter prompts

One surprising part of OpenAI’s model guidance is its recommendation to simplify prompts.

In OpenAI’s internal evaluations, replacing long system prompts with smaller instructions improved scores by roughly 10% to 15%. It also reduced total tokens by 41% to 66% and costs by 33% to 67%.

OpenAI’s explanation is that many behaviors once requiring explicit instructions have become native model behavior. Old prompts often contain years of accumulated warnings, examples, edge cases, and repeated formatting rules.

That extra context can encourage unnecessary exploration and repeated verification.

OpenAI recommends:

Start with the smallest prompt that reliably works.
Expose only the tools relevant to the task.
Keep tool descriptions concise.
Add examples only when testing reveals a specific failure.
Define approval boundaries clearly.
Monitor the context accumulated during a long run.

This does not mean context has stopped mattering. It means useful context should describe the task, evidence, constraints, and expected result rather than micromanaging every mental step. (developers.openai.com)

That fits the workflow we outlined in our complete guide to using AI in 2026: put stable background information inside projects and reusable systems, then keep individual requests focused on the current outcome.

GPT-5.6 versus Claude Fable 5

OpenAI’s release materials paint Sol as a more efficient agentic worker than Claude Fable 5.

The company reports stronger results on Agents’ Last Exam, the Artificial Analysis Coding Agent Index, browsing evaluations, and several tool-heavy professional tasks.

Claude performs better on one prominent coding benchmark. Fable scored 80% on SWE-Bench Pro, compared with Sol’s 64.6% in OpenAI’s published table. (openai.com)

OpenAI published an audit one day before launch arguing that approximately 30% of SWE-Bench Pro’s tasks may be broken. Its reviewers found underspecified prompts, overly strict tests, inadequate test coverage, and prompts that conflicted with hidden requirements.

The company retracted its earlier recommendation that developers rely on the benchmark. (openai.com)

That does not automatically make the weaker score irrelevant.

Developer Simon Willison had early access to Sol and described it as highly capable. He also said it had not yet seemed better than Fable on the complex coding tasks he personally uses.

Axios reported a similar split among early testers. Some viewed GPT-5.6 as more dependable for daily execution, while others preferred Fable for difficult work and deeper judgment.

A useful working distinction has emerged:

Fable may be the stronger planner, writer, or architectural critic.
Sol may be the more persistent implementer and tool user.
Terra may win everyday tasks through a better cost-performance balance.
Luna may win high-volume workflows where the price multiplies quickly.

This is a hypothesis to test, not a permanent ranking.

The model, its tools, the surrounding agent harness, your instructions, and the task environment all affect the output. A one-point benchmark lead may matter less than whether the product can access the correct files, use the required app, and finish the final 20% of the job.

Which model and mode should you use?

A practical starting guide:

Quick questions and ordinary chat

Use Terra at low or medium effort.

You will get a strong result without spending Sol-level resources on every email rewrite or basic explanation.

Important writing, analysis, or strategy

Start with Sol at medium or high effort.

Move to Max or Pro only when the task has clear stakes and you can judge whether the additional work improved the answer.

Routine automation

Use Luna or Terra.

Run the workflow several times, measure failures, and escalate difficult cases to Sol instead of making Sol handle every step.

Complex coding

Use Sol in Codex.

Give it repository access, concrete completion criteria, test requirements, and clear permission boundaries.

Large projects with independent workstreams

Use Ultra.

Ask separate agents to handle research, implementation, validation, and review. Define how their outputs should be combined.

Visual interface work

Use Sol or Terra with a reference image.

GPT-5.6 has stronger design judgment, but concrete examples still beat adjectives such as “modern,” “premium,” or “beautiful.”

High-stakes work

Use the strongest model your budget permits, then add human review.

Higher model capability reduces some mistakes. It does not transfer responsibility for financial, legal, medical, security, or public-facing decisions.

The product strategy underneath the launch

OpenAI has spent years expanding ChatGPT through separate surfaces.

Chat handled conversation. Codex handled software. Atlas handled agentic browsing. Connectors reached external services. Tasks handled recurring work. Canvas and artifacts handled documents and applications.

The new desktop app pulls those surfaces toward one place.

That gives OpenAI three advantages.

First, users can move context between different types of work. Research gathered in the browser can become a presentation, a site, a software task, or a recurring automation.

Second, OpenAI can hide technical complexity from mainstream users without removing it for developers. Work and Codex can share an engine while showing different levels of detail.

Third, the desktop gives the agent access to local files, applications, and workflows that a browser tab cannot reach as easily.

AI competition is moving up the stack. Model intelligence still matters, but the interface, integrations, permissions, memory, scheduling, and execution environment increasingly decide which product people trust with real work.

That is workflow gravity. The more tools, context, projects, and habits a product accumulates, the harder it becomes to replace.

The new app also creates a usability problem

One product now contains more capability. It also asks the user to understand more distinctions.

Should you open Chat, Work, or Codex?

Should you choose Luna, Terra, or Sol?

Should Sol use medium, high, xhigh, or max?

Does the task need standard mode, Pro mode, or Ultra?

Should it use a plugin, the browser, or Computer Use?

The choices make sense to the people building agent systems. They may feel excessive to someone who opened ChatGPT to summarize a meeting.

The strongest version of OpenAI’s strategy eventually hides most of that routing. A user states the goal, the system selects the right model and tools, and the interface only surfaces consequential tradeoffs.

The weaker version gives people a cockpit when they asked for a car.

Launch-day reactions already reflected that tension. Some developers worried that folding Codex into a broad workbench would dilute a focused product. Other users welcomed having chats, repositories, plugins, browser context, and long-running work under one login.

OpenAI must prove that unification reduces friction rather than relocating it.

What to try first

The app contains enough new functionality to turn “exploring” into three hours of clicking through menus.

Start with one real workflow.

1. Update or download the desktop app

The new application is available on Windows and macOS. Existing Codex users can update their current app. Existing ChatGPT users may see the older application renamed ChatGPT Classic.

2. Open Work and choose a familiar task

Pick work where you already know what a good result looks like.

Examples:

Prepare a weekly leadership update.
Analyze a customer-feedback export.
Build a project-status dashboard.
Research five competitors.
Turn meeting notes into a presentation.
Create a small internal calculator.

Familiarity makes failures visible.

3. Connect one relevant plugin

Give the task access to the minimum context it needs.

A weekly update may need Slack and Google Drive. A sales analysis may need Salesforce and a spreadsheet. Avoid connecting every account on the first day.

4. Use Plan mode

Ask Work to inspect the available material and propose a plan.

Review the sources, steps, approval gates, and expected output before execution.

5. Start with Terra

Terra should cover many everyday workflows. Escalate to Sol when Terra misses important context, struggles with judgment, or cannot finish the project reliably.

6. Keep external actions gated

Allow reading, analysis, and reversible local changes. Require approval for messages, purchases, deletions, publishing, permission changes, and sensitive data transfers.

7. Turn a successful workflow into a schedule

Once the result works manually, ask it to repeat weekly or monitor for a specific change.

Automation should come after reliability.

What to watch next

Five signals will reveal whether this becomes a real work platform or another crowded AI interface.

Reliability across long runs

Agents can look impressive for 20 minutes and lose the plot in hour three. Watch whether Work preserves constraints, handles failures, and verifies the final deliverable after long sessions.

Usage limits

Complex Work and Ultra tasks consume more of a plan’s allowance. The product becomes difficult to depend on if teams cannot predict when they will hit a limit.

Approval design

The ideal agent acts independently on safe steps and pauses before consequential ones. Too many prompts destroy the time savings. Too few create risk.

The Atlas transition

The built-in browser and Chrome sidebar must preserve the valuable parts of Atlas without forcing users into another disruptive migration.

Automatic routing

The interface becomes dramatically easier when OpenAI can reliably choose the model, reasoning level, and tools on the user’s behalf.

The launch delivered enough raw capability. The next challenge is making that capability feel obvious.

OpenAI released GPT-5.6 as a family of smarter and more efficient models. The new desktop app reveals the larger ambition: ChatGPT wants to become the operating layer between your goal and the applications required to finish it.

That creates a more useful question than whether Sol beats Fable on this week’s leaderboard.

Will people trust one AI system to hold their context, use their computer, coordinate their agents, and carry an afternoon of work to completion without constant supervision?

The company that earns that trust may own far more than the chat box.