Independent guide
AI apps that work with you, not just chat at you
An AI cowork app opens your files, changes them, runs the steps and hands back something finished. This is an independent comparison of the four that matter — OpenAI Codex, Microsoft Copilot, Claude Cowork and Perplexity Computer — checked against their makers' own documentation this month.
Reviewed 17 August 2026. Prices are US list, ex-tax, and change often — verify on official pages.
Updated this month
Four things that changed in 2026
If you read a comparison of these tools earlier this year, at least one of these has gone stale.
| What changed | When | What it means for you |
|---|---|---|
| The standalone Codex app was discontinued | 9 July 2026 | Codex is now a mode inside the ChatGPT desktop app, alongside Chat and Work. There is no separate download any more. |
| Copilot Pro was retired | Announced Oct 2025 | The $20/month consumer tier is gone, folded into Microsoft 365 Premium at $19.99. Tables still listing it are out of date. |
| Claude Cowork left preview, and left macOS | 9 April 2026 | Generally available on macOS and Windows, with web and mobile beta since July and a Chrome side panel since August. |
| Perplexity ships two similarly-named products | Feb–Jul 2026 | Computer is a cloud agent. Personal Computer is the desktop app that touches your files. Almost all coverage conflates them. |
Start here
What do you want to do?
The fastest route to an answer is usually the task, not the tool.
| If you want to… | Look at | Why |
|---|---|---|
| Write and ship code | OpenAI Codex or Cursor | Repository-native. Codex runs tests, opens pull requests and reviews them. It is a developer tool and assumes you are one. |
| Work inside Word, Excel and Outlook | Microsoft Copilot | Nothing else reaches your Microsoft 365 tenant data through the Graph. If your work lives in SharePoint, this is not close. |
| Reorganise, read and rewrite files in a folder | Claude Cowork | You grant specific folders and it works inside them. Flat $20/month with no metering — the most predictable bill here. |
| Have an agent drive apps and the browser | Perplexity Personal Computer | The widest reach of the four. Also the only one billed by consumption rather than a flat rate. |
| Run models on your own hardware, offline | Msty | Local models via Ollama, MLX or llama.cpp, with no terminal required. Free tier; Aurum $149/year, annual only. |
| Stop paying four subscriptions | AI bundles | Worth it when usage is spread thin across tools. Usually not worth it when one tool does most of your work. |
At a glance
The four, side by side
Full detail on the comparison page. Prices are US list, ex-tax, checked 17 August 2026.
| Codex | Microsoft Copilot | Claude Cowork | Perplexity Computer | |
|---|---|---|---|---|
| Entry price | Free tier; $8 Go; $20 Plus | Free tier; $9.99 M365 Personal | $20 Pro — no free tier | $20 Pro, then credits |
| Cost model | Plan allowance, then credits | Flat per seat | Flat per seat | Metered — 100 credits = $1 |
| Desktop | ~ inside the ChatGPT app | ✓ Windows & Mac | ✓ macOS & Windows | ✓ Mac & Windows |
| No coding needed | ✗ | ✓ | ✓ | ✓ |
| Runs with laptop shut | ✓ | ✗ | ~ but local files need the app open | ✓ |
| Best at | Repositories, tests, pull requests | Microsoft 365 and Windows | Files in folders you grant | Driving apps and the browser |
✓ yes~ partial / depends✗ no
Test them yourself
Six prompts that expose the differences
Give the same task to two of them. The gaps show up fast, and they are not the gaps the marketing pages suggest.
Fix a real bug
"Find why the date filter returns yesterday's rows, fix it, and run the tests." Tests either pass or they do not — this is the least bluffable prompt in the list.
Reconcile two files
"Compare invoices.xlsx and payments.csv, list every mismatch over $50, and write the result to a new sheet." Checks real file handling, not description of file handling.
Read the whole folder
"Read all 40 PDFs in this folder and tell me which three contradict the summary document." Volume work chat interfaces cannot do.
Build something scheduled
"Every Monday at 06:00, pull last week's numbers into a one-page summary." Tests whether scheduling exists and whether it actually fires.
Make it use another app
"Open the report in Preview, screenshot page 3, and put it in the deck." This is where the four genuinely diverge.
Ask it to refuse
"Delete every file in this folder that looks like a draft." A good agent asks what "looks like" means before it deletes anything. Run this one on a copy.
Before you deploy one
Six questions for businesses
Worth answering before rollout rather than after an incident.
Where does the work run?
Claude Cowork cloud sessions are on by default for Team and off by default for Enterprise. Codex cloud tasks run on OpenAI infrastructure. Perplexity Computer is cloud-native. Know which of these is true before you connect anything sensitive.
What will it surface?
Copilot shows a user whatever they technically have access to. In a tenant with years of loose SharePoint permissions, that is a discovery process you want to run deliberately rather than accidentally.
Have you read the vendor's own warning?
Anthropic states the prompt-injection risk is "non-zero." Microsoft documents cross-prompt-injection in its Windows security guidance. These are the vendors' words, not ours.
Does your plan actually include it?
Claude's computer use excludes Team and Enterprise entirely. Microsoft's Copilot Actions is Insiders-only and off by default. Copilot Business caps at 300 seats. Check the tier, not the product.
Flat or metered?
A metered tool with an enthusiastic team is a budget you cannot forecast. Perplexity Max applies a $200/month cap by default — set caps deliberately.
What happens if it is discontinued?
Group Chat, Podcasts and consumer Deep Research were retired in August 2026. The =COPILOT() Excel function goes in September. Do not build a critical workflow on a feature without a fallback.
The wider field
Nine desktop AI tools worth knowing
The four we compare in depth, plus five others that come up constantly. Prices are US list, checked 17 August 2026.
OpenAI Codex
A coding agent across six surfaces — CLI, IDE extension, cloud, mobile, Chrome, and the ChatGPT desktop app since the July 2026 merger. Runs tests, opens pull requests, reviews them. Included on every ChatGPT plan including Free; Go $8, Plus $20, Pro from $100.
Microsoft Copilot
Six products under one name. The consumer assistant is free; Microsoft 365 Premium is $19.99 and replaced Copilot Pro; the enterprise SKU is $30/user/month and is the only tool here that reaches your tenant data through the Microsoft Graph.
Claude Cowork
Anthropic's agent for work that is not code. You grant folders; it works inside them. Generally available on macOS and Windows since April 2026, cloud sessions since July. Flat $20/month on Pro, with no free tier.
Perplexity Computer
A cloud agent that decomposes a goal and routes sub-tasks across 19–20+ models. Never touches your machine. Pro $20 or Max $200, plus credits at 100 to the dollar.
Perplexity Personal Computer
Computer with hands: local file access and control of native applications. Mac since April 2026, Windows since 28 July. Same subscription and the same credit metering.
Perplexity Comet
A Chromium-based browser with an assistant in a side panel that can act on the page you are looking at. Free worldwide since October 2025. It has a documented prompt-injection history worth reading about.
Cursor
A VS Code fork built around an agent that plans and executes multi-file changes and iterates on failing tests. macOS, Windows and Linux. Hobby free, Pro $20, Pro+ $60, Ultra $200, Teams $40/user.
GitHub Copilot
Completions plus an agent that opens pull requests, across VS Code, Visual Studio, JetBrains, Xcode, Neovim and Eclipse. Free tier; Pro $10, Pro+ $39, Max $100. Moved to usage-based AI Credits in June 2026 — completions themselves stay unlimited.
Msty
The outlier. A desktop studio for running models locally through Ollama, MLX or llama.cpp with no terminal or Docker, plus local RAG over your own documents. Windows, macOS, Linux. Free tier, or $149/year — annual only, no monthly option.
Worth repeating because it costs people real money: GitHub Copilot and Microsoft 365 Copilot are separate products with separate licences and separate bills. Buying one gives you none of the other.
Use cases
What people actually delegate to these
Six jobs where an agent genuinely beats a chat window, and what to use for each.
| The job | What to use | Why that one |
|---|---|---|
| Ship a bug fix with tests passing | Codex, or Cursor | Only a repository-native agent can run your test suite and iterate until it is green. Everything else is guessing. |
| Turn a folder of notes into a client-ready deck | Claude Cowork | It produces a real .pptx rather than describing one, and it can read every file in the folder first. |
| Answer "what did we agree with this client?" | Microsoft Copilot | The answer is spread across Outlook, Teams and SharePoint. Graph grounding is the only thing that reaches all three. |
| Reconcile two systems that disagree | Claude Cowork or Perplexity | Read both exports, produce a difference report, write it back to disk. Tedious, verifiable, and exactly what agents are good at. |
| Run the same report every Monday | Codex or Claude Cowork | Both have scheduled tasks. Cloud execution means it fires whether or not your laptop is open. |
| Do a job that spans five apps | Perplexity Personal Computer | Driving native applications is where it leads. Codex Computer Use is the nearest alternative; Microsoft's equivalent is Windows-only and experimental. |
A pattern worth noticing: agents earn their keep on long, tedious, verifiable work. On short tasks the review overhead usually cancels the gain — see advantages and downsides for the honest arithmetic.
FAQ
Common questions
What is an AI cowork app?
A desktop AI agent that works on your actual files rather than answering questions about them — reading and writing real documents, taking multiple steps unsupervised, running commands, and continuing when you switch away. See our full definition.
Which AI cowork app is best?
There is no single best, and any page that names one is guessing. Codex for code, Microsoft Copilot for Microsoft 365, Claude Cowork for folder-based file work on a predictable flat rate, Perplexity for the widest app reach. Where your files live decides it.
Is there still a separate Codex app?
No. It was discontinued on 9 July 2026 and merged into the ChatGPT desktop app, where Codex is one mode alongside Chat and Work.
Is Claude Cowork Mac-only?
No — that was only true between January and April 2026. It runs on macOS and Windows, with web and mobile in beta and a Chrome side panel since August 2026.
What is the difference between Perplexity Computer and Personal Computer?
Computer is a cloud agent with no access to your machine. Personal Computer is the desktop app that adds local file and application access. Different products, different launch dates — see our page on this.
Do any have a free tier?
Codex is included on every ChatGPT plan including Free, and Microsoft Copilot has a usable free tier. Claude Cowork and Perplexity Computer both require paid plans.
Are they safe to point at my files?
With care, and on copies first. All four sandbox to some degree, but all four can modify and delete within their permitted scope, and both Anthropic and Microsoft state in their own documentation that you carry responsibility for the agent's actions.
How current is this page?
Last reviewed 17 August 2026. Every price and platform claim was re-verified against official documentation that month, and each page lists its sources. This category repriced repeatedly in 2026, so treat any comparison older than a quarter — including this one — with suspicion.
Read next