Originally published at devtoolpicks.com
On April 16, one hour after Anthropic launched Claude Opus 4.7, OpenAI shipped a Codex update that went way beyond model weights. Codex now sees your Mac screen, clicks around in native apps, runs 90+ plugins, generates images inline, and remembers your preferences across sessions. OpenAI is no longer pitching Codex as a terminal coding tool. It is pitching it as a full desktop AI agent that happens to do code really well.
The tweet that summed up the day had 1.6M views and read: "OpenAI just stole the spotlight one hour after Opus 4.7 dropped. Anthropic bets on raw model power, OpenAI wants your whole workflow." That is the story.
So if you are currently using Claude Code and watching your X feed light up with Codex demos, the question is: should you switch? Here is an honest breakdown of what actually changed, where each tool still wins, and what it means for indie hackers and solo devs.
Quick Verdict
| Question | Answer |
|---|---|
| Biggest Codex change today | Computer Use on macOS, 90+ plugins, persistent memory |
| Best for terminal-native agentic coding | Claude Code (still) |
| Best for desktop automation across apps | Codex (new) |
| Cheapest entry tier | Both $20/mo (ChatGPT Plus vs Claude Pro) |
| Best $100 tier value for heavy use | Close call, see breakdown below |
| EU/UK availability | Claude Code: yes. Codex Computer Use: no at launch |
| Worth switching if you are happy with Claude Code | Probably not yet |
What Actually Changed in Codex Today
OpenAI announced the update on April 16, 2026, and it is the biggest Codex release since the Mac app launched in February. Here is what shipped.
Computer Use on macOS. Codex can now operate Mac apps with its own cursor. It sees what is on your screen, clicks buttons, types into fields, and completes tasks in native apps that do not expose APIs. OpenAI emphasizes this runs in the background so multiple Codex agents can operate your Mac in parallel without interfering with your own work. The feature is built by the team OpenAI acquired from Sky Applications (the original Apple Shortcuts/Workflow team). It is not available in the EEA, UK, or Switzerland at launch.
In-app browser. A built-in browser powered by OpenAI's Atlas technology. You can open local files or public pages that do not require sign-in, comment directly on the rendered page, and ask Codex to act on your feedback. OpenAI plans to expand this into full browser control over time.
Image generation. The gpt-image-1.5 model is integrated directly into Codex. No more switching to ChatGPT to generate UI mockups, diagrams, or design assets.
Persistent memory. Codex remembers preferences, workflows, tech stacks, and corrections across sessions. So you stop re-explaining your stack every morning.
90+ new plugins. Jira (via Atlassian Rovo), CircleCI, GitLab Issues, Microsoft Suite, Neon by Databricks, Slack, Notion, Google Workspace, and MCP servers. OpenAI says the curated plugin approach is a deliberate response to security concerns in more open agent platforms.
Long-running automations. Codex can reuse existing conversation threads, schedule future work for itself, and wake automatically to continue tasks across days or weeks.
Other improvements. Multiple terminal tabs, SSH into remote devboxes (in alpha), rich file previews in a sidebar for PDFs, spreadsheets, slides, and docs, a summary pane for plans and artifacts, GitHub review comment addressing, and first-time support for Intel Macs.
Under the hood, GPT-5.4 is the recommended Codex model with an experimental 1 million token context window. OpenAI says 3 million developers now use Codex weekly.
Computer Use: The One Feature Claude Code Does Not Have
This is the single biggest functional difference after today.
Claude Code is terminal-native. It operates on your codebase through the CLI, reads and writes files, runs shell commands, and works with any editor. It does not click around in GUI apps.
Codex now does both. It operates your codebase through the CLI just like Claude Code, and it can also drive native Mac apps: Xcode simulators, design tools, browser-based internal dashboards, app testing flows, anything that does not expose an API.
For most backend and web development, this does not matter. You are in your terminal and your editor, you are not clicking around native apps. Claude Code remains equal or better there.
For some specific workflows, Computer Use is genuinely a step-change:
- Native iOS and macOS development. Testing flows in Xcode simulators, capturing screenshots, iterating on Swift UI changes by actually clicking the running app
- Non-API-exposed tools. Internal dashboards, legacy apps, or admin panels where you would otherwise be writing Selenium scripts
- Design iteration. Pushing code changes, opening the rendered page in the in-app browser, commenting on what needs to change, letting the agent iterate
If you do any of these daily, Codex is now the clear winner on capability. If you do not, the feature is mostly marketing.
Codex vs Claude Code: Head to Head
Developer experience
Claude Code is a CLI tool that runs in any terminal. You stay in your editor of choice (VS Code, JetBrains, Zed, Neovim, whatever). It reads your codebase, makes multi-file edits, and runs shell commands. The IDE integrations for VS Code and JetBrains add side-panel conversations without replacing your editor.
Codex now has three surfaces: the CLI, the IDE extensions, and the Codex macOS app. The Mac app is where the new features live. It includes worktrees so multiple agents can work the same repo in parallel without conflicts. If you work in a single terminal or single editor window today, the Codex app is a different mental model from Claude Code. Not worse, just different.
Winner: Tie. Claude Code wins if you want editor-agnostic terminal simplicity. Codex wins if you want a dedicated agent workspace with built-in multi-agent orchestration.
Model quality
GPT-5.4 (Codex default) has an experimental 1M token context window. Claude Opus 4.7 launched yesterday with its own benchmark gains but a 200K standard context window on consumer plans (500K on Enterprise Claude).
For raw coding benchmarks, both are close enough that real-world preference often comes down to individual tasks. Anthropic has historically led on multi-file refactors and deep codebase understanding. OpenAI has historically led on speed and execution. Opus 4.7 and GPT-5.4 both claim improvements in those respective areas. We reviewed Opus 4.7 specifically here if you want the benchmark breakdown.
Winner: Roughly tied. Task-dependent. Run your own test on a real task in your codebase before switching.
Pricing
Both companies ended up at the same pricing structure, which is not a coincidence.
| Tier | OpenAI (Codex included) | Anthropic (Claude Code included) |
|---|---|---|
| Entry paid | Plus: $20/mo | Pro: $20/mo ($17 annual) |
| Middle | Pro: $100/mo (5x Plus limits) | Max 5x: $100/mo (5x Pro limits) |
| Top | Pro: $200/mo (20x Plus limits) | Max 20x: $200/mo (20x Pro limits) |
OpenAI launched the $100 Pro tier on April 9 explicitly to match Anthropic's Max 5x. Through May 31, 2026, the new $100 Codex plan gets a temporary 10x Plus boost instead of 5x.
The price-per-tier is identical. The usage you get inside each tier depends on how heavy your daily sessions are and which tool's rate limits hit you harder. Reports from heavy Claude Code users on Max 20x still hitting rate limits on long days are common. Codex users report similar friction on Plus. For indie hackers running 4-6 hours of agent work daily, neither $20 tier is actually enough and both $100 tiers feel tight.
Winner: Tie on paper. On real usage, both leak. Pick based on which model you prefer, not which is cheaper.
Ecosystem and integrations
Codex today shipped 90+ plugins, bringing it above 100 total when combined with existing ones. The curated ecosystem includes Atlassian Rovo for Jira, CircleCI, GitLab, Slack, Notion, Microsoft Suite, Google Workspace, Neon, and MCP servers.
Claude Code has MCP server support and a growing ecosystem through Anthropic's marketplace launched in March 2026 (with Snowflake, Harvey, Replit as launch partners). The MCP protocol is open and has more community-built servers than OpenAI's curated plugin list.
Winner: Codex wins on curated breadth for today's announcement. Claude Code wins on the open protocol (MCP) where developers can self-build integrations.
Regional availability
Claude Code works in the EU, UK, Switzerland, and everywhere else Anthropic operates.
Codex Computer Use is explicitly not available in the EEA, UK, or Switzerland at launch. Codex personalization and memory features are still rolling out for Enterprise, Edu, and users in the EU and UK.
If you are an indie hacker in Europe (or a Turin-based developer like me), Codex Computer Use is off the table entirely right now.
Winner: Claude Code, decisively, if you are in the EU or UK.
Who Should Actually Switch
Switch to Codex if:
- You do native macOS or iOS development and the Computer Use + Xcode simulator workflow is a clear upgrade
- You want a dedicated multi-agent workspace rather than a terminal-based approach
- You already live inside the OpenAI ecosystem (ChatGPT Plus subscribers who use it heavily)
- You need the 90+ plugins out of the box and do not want to self-configure MCP servers
Stay on Claude Code if:
- You are in the EU, UK, or Switzerland (Computer Use is blocked, which kills the main differentiator)
- Your workflow is web/backend development inside a terminal and you do not need desktop automation
- You prefer the open MCP protocol over curated plugins
- You are on a Max 20x plan and getting real value from the higher Claude Code rate limits
- You value model quality on long multi-file refactors and believe Anthropic's models still edge ahead there
Test both if:
- You are a solo developer or indie hacker with no strong existing preference
- You have $20/mo to spare for a month of testing the other platform
- You are about to start a new project and can evaluate both against the same real task
Honest Cons of the New Codex
Every Codex launch review glossed over the cons. Here are five worth naming.
Mac-only Computer Use. Windows users get the Codex app with plugins and everything else, but not Computer Use. Linux users get none of it. If your main machine is not a Mac, the headline feature of today's launch is not for you.
Regional lockouts. Computer Use is blocked in the EEA, UK, and Switzerland. Memory and personalization are also staggered for those regions. Half of Europe cannot use the announcement's biggest features today.
Platform lock-in risk. OpenAI's superapp strategy (per WSJ reporting) is to unify ChatGPT, Codex, and Atlas browser into one product. That means deeper OpenAI ecosystem lock-in over time. If you want tool-agnostic flexibility, terminal-native Claude Code has less of this problem.
The pricing rebalance. On April 9, OpenAI "rebalanced" the Plus plan's Codex allocation to "support more sessions throughout the week instead of longer sessions in a single day." Translation: if you were running heavy Codex sessions on $20 Plus, you may now hit limits faster and be nudged toward the $100 tier. Claude Code has done similar rebalancing at $20 historically. Neither company is above this move.
Curated plugin ecosystem. OpenAI frames the curated approach as a security win, which is reasonable. It also means if a plugin you need is not in the 90+ list, you wait for OpenAI to add it or you use MCP (which Codex does support). Claude Code's MCP-first approach is more open at the cost of some security guardrails.
Pricing Table
| Plan | Monthly | Claude Code usage | Codex usage | Notes |
|---|---|---|---|---|
| ChatGPT Plus | $20 | N/A | 1x baseline | OpenAI side of $20 tier |
| Claude Pro | $20 ($17 annual) | Full Claude Code access | N/A | Anthropic side of $20 tier |
| ChatGPT Pro $100 | $100 | N/A | 5x Plus (10x until May 31) | OpenAI mid-tier |
| Claude Max 5x | $100 | 5x Pro limits | N/A | Anthropic mid-tier |
| ChatGPT Pro $200 | $200 | N/A | 20x Plus | OpenAI top tier |
| Claude Max 20x | $200 | 20x Pro limits, priority access | N/A | Anthropic top tier |
| API pay-per-token | varies | Sonnet 4.6: $3/$15 per 1M, Opus 4.7: $5/$25 per 1M | GPT-5.4: varies by token type | Best for unpredictable usage |
Both ecosystems top out at $200/mo for consumer plans. Above that, you are in Business or Enterprise territory with custom pricing.
FAQ
Is Codex now better than Claude Code overall?
No. It is better at new things Claude Code does not do (Computer Use on Mac, curated plugin ecosystem, in-app browser, image gen). Claude Code remains competitive or better at terminal-native agentic coding, multi-file refactors, and EU/UK accessibility. The right answer depends on your workflow.
Can I use Codex Computer Use in the EU?
Not at launch. OpenAI explicitly blocks Computer Use in the European Economic Area, the UK, and Switzerland. The rest of Codex (the app, the CLI, most plugins) works in those regions, but the headline feature from today's launch is geofenced.
Does Codex have MCP support?
Yes. Codex supports MCP servers alongside its curated plugin ecosystem. The difference is curation: OpenAI vets the 90+ plugins in its directory, while MCP servers are community-built and self-configured.
Which is cheaper for heavy users?
At the $100 and $200 tiers, they are priced identically. Real usage varies by tool. Claude Code users on Max 20x report hitting rate limits on long days. Codex users on Plus report the same. Neither is meaningfully cheaper in real-world heavy use.
Should I cancel my Claude Code subscription and switch to Codex?
Probably not today. The EU/UK availability issue, the Mac-only Computer Use limitation, and the fact that Claude Code is still strong at its core use case (terminal-native agentic coding) all argue for staying put unless one of the new Codex features specifically changes your workflow. We compared all three major AI coding subscriptions here if you want the full context.
What about Cursor 3?
Cursor 3 shipped parallel agents and a redesigned composer window on April 2, 2026. It competes with both Codex and Claude Code on the IDE-integrated side of the market. If you want a polished IDE-first experience rather than a CLI or separate agent app, Cursor 3 is still a strong option. It does not do Computer Use like Codex now does.
What about Claude Code's new desktop app?
Anthropic redesigned the Claude Code desktop app with parallel sessions and an integrated terminal earlier this month. We reviewed that redesign here. It narrows the UX gap with the Codex app but does not match Codex Computer Use.
Final Recommendation
For most indie hackers and solo developers, the honest answer is do not switch yet.
If you are an EU or UK developer, the Computer Use lockout alone kills the main reason to move. If you are on a Mac in the US doing web/backend work, the features Codex gained today are nice but do not meaningfully change your day-to-day workflow. If you are a native iOS or macOS developer, Computer Use genuinely opens up workflows that were not practical before and Codex deserves a test.
What today's announcement actually signals is bigger than one product update. OpenAI is done competing on just model quality. They are building Codex into a desktop AI superapp that wraps your whole workflow. Anthropic is still competing on model intelligence and terminal-native agentic coding. Both bets are reasonable. Neither is wrong.
The boring truth: try Codex for a week if you have the budget. Keep Claude Code running alongside it. After seven days, you will know which one fits your actual workflow, not the demo workflow.
Opus 4.7 and Codex Computer Use launching the same day is the strongest signal yet that the AI coding war is fully on. The winners from here are not the subscribers picking the right horse. The winners are the indie hackers shipping faster with whichever tool fits best.
Top comments (0)