Full comparison
Codex vs Copilot vs Claude Cowork vs Perplexity Computer
Four desktop AI agents, checked against their makers' own documentation in August 2026. What each one is, who it suits, what it actually costs once metering is included, and where each falls down. There is no single winner, and any page that gives you one is guessing.
Reviewed 17 August 2026. Prices are US list, ex-tax, and change often — verify on official pages.
Start here
Four things changed since most comparisons were written
If you have read another comparison of these tools this year, at least one of these has probably gone stale on you.
| What changed | When | Why it matters |
|---|---|---|
| The standalone Codex desktop app no longer exists | 9 July 2026 | OpenAI merged it into the main ChatGPT desktop app, where Codex is now one mode alongside Chat and Work. If you are looking for a separate Codex download, you will not find one. |
| Copilot Pro was retired | Announced October 2025 | The $20/month standalone consumer tier was folded into Microsoft 365 Premium at $19.99/month. Any table still listing "Copilot Pro" is out of date. |
| Claude Cowork left research preview and left macOS | 9 April 2026 | It is now generally available on macOS and Windows, with web and mobile in beta since July and a Chrome side panel since August. "Mac-only research preview" has been wrong for four months. |
| Perplexity ships two different products with confusingly similar names | Feb–Jul 2026 | Computer is a cloud agent. Personal Computer is the desktop app that touches your local files. Most coverage conflates them. See our Perplexity Computer page. |
Feature matrix
Thirteen things that actually differ
Green does not mean "the winner" — it marks where a capability is fully present rather than partial or absent.
| Feature | OpenAI Codex | Microsoft Copilot | Claude Cowork | Perplexity Computer |
|---|---|---|---|---|
| Maker | OpenAI | Microsoft | Anthropic | Perplexity |
| Desktop app | ~ inside the ChatGPT app | ✓ Windows & Mac | ✓ macOS & Windows | ✓ Mac & Windows (as Personal Computer) |
| Linux | ~ preview since Aug 2026 | ✗ | ~ download exists, computer use unavailable | ✗ |
| Runs with the laptop closed | ✓ cloud tasks | ✗ | ✓ cloud sessions since Jul 2026 | ✓ cloud-native |
| No coding required | ✗ developer tool | ✓ | ✓ | ✓ |
| Works on a local folder you choose | ~ repo / workspace sandbox | ~ Documents, Downloads, Desktop, Pictures — experimental | ✓ you grant specific folders | ✓ via Personal Computer |
| Autonomous multi-step agent | ✓ | ~ assistant plus separate agents | ✓ | ✓ |
| Built for coding | ✓ strongest here | ~ that is GitHub Copilot, sold separately | ~ capable, not specialised | ~ can produce code |
| Office documents (Word, Excel, PowerPoint) | ✗ | ✓ native, in-app | ✓ creates real .xlsx and .pptx files | ~ produces documents |
| Controls other apps on your screen | ✓ Computer Use, macOS & Windows | ~ Copilot Actions — Windows, experimental, off by default | ~ research preview, Pro/Max only | ✓ core to Personal Computer |
| Terminal / IDE integration | ✓ CLI, VS Code, Cursor, Windsurf, Xcode, JetBrains | ~ via GitHub Copilot | ✗ | ✗ |
| Cost model | Plan allowance, then credits | Flat per seat | Flat per seat | ✗ credit-metered, pay as you go |
| Entry price (individual) | Free tier; $8 Go; $20 Plus | Free tier; $9.99 M365 Personal | $20 Pro (no free tier) | $20 Pro, then credits |
✓ yes~ partial / depends✗ no
A note on that last row: three of these four are flat-rate at the entry tier. Perplexity is the exception — the subscription buys access, and the work itself draws on credits at 100 credits = $1. That is the single biggest structural difference in this table.
Head to head
The match-ups people search for
Short verdicts, with the reasoning rather than just the conclusion.
| Match-up | Verdict |
|---|---|
| Codex vs Claude Cowork | Codex is a coding agent that happens to run on your desktop; Cowork is a general work agent that happens to be able to code. If your task has a repository, Codex. If your task has a folder of documents, Cowork. Anthropic itself describes them as the same engine with a different interface — Claude Code for engineering, Cowork for everything else. |
| Copilot vs Claude Cowork | Different lanes, and the deciding factor is where your files live. If your work is in SharePoint, Outlook and Teams, Copilot reaches data nothing else can touch, because it is grounded in the Microsoft Graph. If your work is loose files on your own disk, Cowork is the more autonomous agent. Copilot Vision, note, only looks — it does not click or type for you. |
| Codex vs Microsoft Copilot | Barely comparable. Codex is an autonomous coding agent; Microsoft's coding product is GitHub Copilot, which is a separate licence and a separate bill. For Office work Copilot wins by default. For shipping code, Codex is the more autonomous of the two. |
| Perplexity Computer vs Claude Cowork | The closest genuine match-up on this page, and the one nobody covers properly. Both take a goal and work across your files. Cowork is flat-rate and sandboxed to folders you grant. Perplexity's Personal Computer is broader — it drives native apps and the browser — but it is credit-metered, so a heavy month costs materially more than $20. |
| All four, for non-coders | Codex is out; it is a developer tool. That leaves Copilot if you live in Microsoft 365, Cowork if you want a predictable bill and folder-level control, and Perplexity if you want the widest reach across apps and can tolerate a variable cost. |
Pricing
What each one really costs
US list prices in USD, excluding tax, checked 17 August 2026. Swiss and EU pricing differs. The "watch out" column is where the headline number stops telling the whole story.
| Product | Individual entry | Higher tiers | Watch out for |
|---|---|---|---|
| OpenAI Codex | Included on every ChatGPT plan, including Free. Go is $8/mo, Plus $20/mo. | Pro from $100/mo (5× Plus usage), $200/mo (20×). Business $20/user/mo on annual, $25 monthly. | Codex draws from the same usage pool as ChatGPT Work, ChatGPT for Excel and Workspace Agents — non-coding use eats your coding capacity. Beyond the plan allowance you buy credits, which expire after 12 months. OpenAI does not publish limits for Free or Go. |
| Microsoft Copilot | Free tier. Microsoft 365 Personal $9.99/mo, Family $12.99/mo, Premium $19.99/mo (the old Copilot Pro). | Microsoft 365 Copilot for enterprise $30/user/mo on annual commitment ($31.50 monthly). Copilot Business, capped at 300 seats, $21 list — currently $18 on promotion. | On Family and Premium the AI benefits go to the subscription owner only, not the other household members. The $18 Business rate is promotional and ends 30 September 2026. The =COPILOT() Excel function is being retired on 14 September 2026. |
| Claude Cowork | No free tier. Pro $20/mo, or $17/mo billed annually. | Max 5× $100/mo, Max 20× $200/mo. Team $25/seat/mo ($20 annual). Enterprise $20/seat/mo plus usage at API rates. | The most predictable bill of the four — flat rate, no credits. The trade-off is that usage limits are time-windowed and have been a persistent complaint; Anthropic ran two temporary limit-doubling promotions in 2026, both now ended. |
| Perplexity Computer | Pro $20/mo or Max $200/mo. Not on the free tier. | Enterprise Pro and Enterprise Max, with larger monthly credit allocations. | Credit-metered, not flat-rate. 100 credits = $1. Max includes 10,000 credits/month; Pro includes no monthly allocation, only a one-time 4,000-credit bonus that expires in 30 days. A single large project can consume thousands of credits. Max applies a $200/month spend cap by default, adjustable up to $5,000. |
If predictable cost matters more than raw capability, that table has an obvious answer, and it is not the most capable tool. Perplexity's Personal Computer does the most; it is also the only one where a busy month can produce a bill you did not plan for.
If the job is code
Codex is not the only coding agent
Two names come up constantly in this category and are worth knowing before you settle on Codex.
Cursor
A fork of VS Code, so your extensions, themes and keybindings carry across. The agent plans and executes multi-file changes, runs terminal commands and iterates on failing tests. macOS, Windows and Linux, plus a CLI and cloud agents. Hobby free, Pro $20/mo, Pro+ $60, Ultra $200, Teams $40/user. Usage is limit-metered, and overage is the most common complaint.
GitHub Copilot
Completions plus an agent that opens pull requests, in VS Code, Visual Studio, JetBrains, Xcode, Neovim and Eclipse. Free tier; Pro $10/mo, Pro+ $39, Max $100, Business $19/seat, Enterprise $39/seat. Since 1 June 2026 it bills on usage-based AI Credits at $0.01 per credit — though code completions themselves remain unlimited and never consume credits on a paid plan.
Msty
The outlier: a desktop studio for running models locally via Ollama, MLX or llama.cpp, with no terminal or Docker required, plus local RAG over your own documents. Windows, macOS and Linux. Free tier, or Aurum at $149/year — annual only, with no monthly option. Quality and speed are bounded entirely by your own hardware.
Worth knowing: GitHub Copilot and Microsoft 365 Copilot are entirely separate products with separate licences and separate bills. Buying one gives you nothing of the other — a genuinely common and expensive misunderstanding.
How to choose
Answer these five questions instead
Feature matrices are useful but they do not decide anything. These do.
Where do your files actually live?
Microsoft 365 and SharePoint points to Copilot, which reaches tenant data through the Graph that nothing else can see. Loose files on your own disk points to Claude Cowork or Perplexity. A git repository points to Codex or Cursor.
Do you need it to run while you sleep?
Codex cloud tasks, Claude Cowork cloud sessions and Perplexity Computer all continue with your laptop shut. Microsoft Copilot does not. If you want a scheduled Monday-morning report, that rules one option out immediately.
Can you tolerate a variable bill?
Three of these are flat-rate at entry. Perplexity is metered. If your finance process cannot absorb a month that costs three times the last one, that is a decision, not a detail.
Does it need to touch apps, not just files?
Driving native applications is where these genuinely diverge. Perplexity's Personal Computer is built for it. Codex Computer Use does it on macOS and Windows. Claude's version is a research preview limited to Pro and Max. Microsoft's Copilot Actions is Windows-only, experimental, and off by default.
Who carries the risk if it gets something wrong?
You do, in all four cases. Anthropic states plainly that users are responsible for every action Claude takes, including financial transactions. Microsoft documents cross-prompt-injection risk in its own Windows security guidance. Start with reversible work on non-critical files.
The honest part
What each one is genuinely bad at
Every vendor page lists strengths. These are the limitations, taken from the makers' own documentation and from named reporting.
| Product | The real limitation |
|---|---|
| OpenAI Codex | The shared usage pool is the sharp edge — your Codex capacity is consumed by unrelated ChatGPT Work and Excel usage. Network access is off by default in every sandbox mode, which surprises people whose builds need to fetch dependencies. Linux is preview-quality only. And OpenAI publishes no dollar-per-credit figure, so the true cost of heavy use is hard to model in advance. |
| Microsoft Copilot | Adoption is the story here: fewer than 4.5% of Microsoft's 450 million commercial Microsoft 365 subscribers pay for Copilot, and consumer Copilot has flatlined around 20 million weekly users. Feature churn is real — Group Chat, Podcasts and consumer Deep Research were all retired in August 2026, and the =COPILOT() Excel function follows in September. Copilot Vision cannot act; it only observes. |
| Claude Cowork | Cloud sessions can only reach your local files while the desktop app is open and connected, which undercuts the "works with your laptop closed" pitch in exactly the case where you wanted it. Computer use is still a research preview and excludes Team and Enterprise entirely. There is no cross-chat memory in Cowork sessions and no session sharing. Anthropic states its own prompt-injection risk is "non-zero". |
| Perplexity Computer | Cost opacity combined with blast radius. It is the most capable agent here and the least predictable to budget for, and the version that touches your local files also has write access to them and to your email. Perplexity publishes no RAM or CPU minimums for Personal Computer. Comet, the browser it uses as its hands, has a documented prompt-injection history. |
FAQ
Comparison FAQ
Which AI cowork app is best?
There is no single best. Codex for code, Microsoft Copilot if your work lives in Microsoft 365, Claude Cowork for general file work on a predictable flat rate, and Perplexity Computer if you want the widest reach across apps and can tolerate credit-based billing. The deciding factor is almost always where your files already live, not which model is cleverest.
Do any of them have a free tier?
Codex does — OpenAI includes it on every ChatGPT plan including Free, though cloud tasks appear to require Plus or above. Microsoft Copilot has a genuinely useful free tier. Claude Cowork does not; it requires a paid Claude plan starting at $20/month. Perplexity Computer requires Pro or Max.
Is Claude Cowork still Mac-only?
No. That was true only between 12 January and 9 April 2026. It is now generally available on macOS and Windows, with web and mobile in beta since July 2026 and a Chrome side panel since August 2026.
Is there still a separate Codex desktop app?
No. On 9 July 2026 OpenAI discontinued the standalone Codex app and merged it into the main ChatGPT desktop app, where Codex is one mode alongside Chat and Work. Existing users were transitioned automatically with projects and settings preserved.
What happened to Copilot Pro?
It was retired and folded into Microsoft 365 Premium at $19.99/month, which includes the Office apps, Copilot across them, the Researcher and Analyst agents and 1 TB of storage. On Premium and Family plans the AI benefits apply to the subscription owner only.
What is the difference between Perplexity Computer and Personal Computer?
Computer is the cloud agent — it runs entirely on Perplexity's servers with no access to your machine. Personal Computer is the desktop app that adds local file access and control of native applications. They are different products with different launch dates, and most coverage conflates them.
Do I need a Mac mini for Perplexity?
No. That claim comes from a Perplexity changelog entry suggesting a dedicated Mac mini as a convenient way to run an always-on agent. Personal Computer runs on any Mac, and on Windows 10 or 11 since July 2026.
Can these agents delete my files?
Yes, within whatever scope you grant them. All four use sandboxing or approval gates, but all four can also modify and remove files inside their permitted area. Anthropic and Microsoft both state in their own documentation that the user carries responsibility for the agent's actions. Start on a copy of anything you cannot afford to lose.
Is GitHub Copilot the same as Microsoft 365 Copilot?
No, and this is an expensive misunderstanding. They are separate products, separate licences and separate bills. GitHub Copilot is a developer tool grounded in your repositories; Microsoft 365 Copilot is an information-worker tool grounded in your Microsoft Graph. Buying either gives you nothing of the other.
How often do these prices change?
Constantly. Between February and August 2026 alone: Copilot Pro was retired, GitHub Copilot moved to usage-based billing, Microsoft raised suite prices, OpenAI restructured Codex around credits and cut GPT-5.6 prices, and Perplexity extended Computer from Max-only to Pro. Treat any comparison page over three months old as unreliable, including this one — the review date is at the top for that reason.
Sources
Where these facts came from
Every price and capability on this page traces to an official page or named press report. Checked 17 August 2026.