Claude Code hits $200/mo. Codex Desktop runs free

OpenAI's Codex Desktop app quietly outscored Claude Code on the industry's toughest coding benchmark last week — and unlike Claude Code's $200/month price tag, Codex starts completely free. That gap is now the defining story of the AI coding tools market in 2026: the most expensive paid tool just met a serious free competitor that beats it on speed and outscores it on accuracy.

The Pricing Gap That Changes the Math

Anthropic's Claude Code (the terminal-based AI coding agent built into VS Code) charges up to $200/month on the Max plan for heavy users. OpenAI structured Codex entirely differently:

• Free tier — Available on the $0 ChatGPT account with usage throttling; enough for personal projects and testing the full workflow
• Plus plan ($20/month) — Standard ChatGPT Plus includes Codex with doubled rate limits compared to free
• Pro 5X plan ($100/month) — 5× the task allowance of Plus; during the current promotional window (ends May 31, 2026) this becomes an effective 10× allowance — more AI compute per dollar than any competing plan

The math is stark: even Codex's premium plan costs half of Claude Code's ceiling. And during the current promotion, heavy users get twice the compute for half the price. OpenAI has historically kept its most capable tools behind high price points. Making Codex available — even throttled — to $0 accounts is a deliberate land-grab in the developer tools market.

If you've been evaluating AI coding tools, this is the lowest-friction moment in this product category's history to run a real comparison without spending anything.

Head-to-Head: Codex vs Claude Code by the Numbers

The benchmark that matters most here is SWE-Bench Pro (a standardized test that measures how well an AI agent can resolve real, unmodified GitHub software issues without any human help — the closest proxy for actual engineering output). Codex now leads:

Metric	OpenAI Codex	Claude Code
SWE-Bench Pro score	56.8%	55.4%
Output speed (tokens per second — word-chunks generated per second)	1,000+	~200
Context window (how much text the AI holds in memory at once)	200K tokens	1M tokens
Parallel agents (multiple AI instances working simultaneously)	Native support	Limited
Starting price	Free	$200/month (Max plan)

The 5× speed advantage is felt immediately. At 1,000+ tokens per second, Codex writes a complete function while you're still reading Claude Code's first line of output. For tight deadlines or large-scale refactors, this difference is tangible. Claude Code's advantage — a 1M token context window (how much text the AI can hold in mind simultaneously, like working memory) — matters most when you need the AI to reason across an entire large codebase at once. For most single-file or module-level work, Codex's 200K is more than sufficient.

What Codex Desktop Actually Does on Your Computer

Codex Desktop is not a chatbot or an IDE plugin. It's an Electron-based agent platform (a desktop application built with the same framework as VS Code, running natively on your machine) with four capabilities that go well beyond autocomplete:

Parallel Agent Execution via Git Worktrees

Codex spawns multiple AI agents, each working in an isolated git worktree (a separate copy of your project folder that can't overwrite another agent's changes). One agent refactors your authentication module. Another writes unit tests for the API layer. A third updates the README. All simultaneously, with no merge conflicts. Theo Browne, a developer who switched from Cursor to test Codex, described the experience as "fundamentally different" orchestration — not just faster, but architecturally unlike anything else on the market.

Automations — Set Tasks to Run Overnight

You can schedule Codex on a cron schedule (an automated timer — like setting a recurring alarm for "every Monday at 6 AM, summarize last week's CI failures"). Nightly security scans, weekly dependency audits, daily changelog generation, automated issue triage — results land in a review queue for you to approve or dismiss. The AI does the work; you review the output on your own time.

Skills — Teach It Your Team's Conventions

Skills are reusable SKILL.md files (plain-text instruction files your team writes once and checks into version control alongside your code). They encode recurring workflows: how to process Figma design exports, which naming conventions to follow, how to triage Linear tickets by priority. The AI stops guessing at your preferences; it follows documented rules shared across your entire team.

Security Scanning — 792 Critical Bugs Found in Testing

In pre-release testing across public GitHub repositories, Codex Security uncovered 792 critical vulnerabilities and 10,561 high-severity issues — the kind of bugs that survive years of human code review. Every proposed fix is validated in a sandbox (an isolated environment where code executes without touching your live system) before being surfaced. False positives dropped 50% since the security module launched.

The April 16 Update: What Version 0.121.0 Changes

The update ZDNet flagged on April 16, 2026 (version 0.121.0) is the most substantial Codex Desktop release since the Windows launch on March 4. Here's what changed:

# New in v0.121.0: Plugin marketplace install command
codex marketplace add github:your-org/your-plugin
codex marketplace add https://git.example.com/plugin.git
codex marketplace add ./local-plugin-directory
codex marketplace add marketplace.json  # Direct marketplace file

• Marketplace installs — Add plugins from GitHub, git URLs, local directories, or direct marketplace.json files. This single feature could replicate VS Code's extension ecosystem dynamics inside an AI agent.
• Ctrl+R prompt history search — Search your past prompts using reverse search, like searching bash history. Small quality-of-life fix; large time savings for repetitive workflows.
• Memory controls — Manage, reset, or delete Codex's accumulated memory of your projects, including cleanup of stale context that was causing outdated suggestions in older threads.
• MCP expansion — Broader support for the Model Context Protocol (the open standard that lets AI tools connect to external services like databases, local files, and third-party APIs), now with namespaced registration and parallel tool calls — meaning Codex can query multiple services at the same time.
• Realtime API additions — Output modality controls, transcript completion events, and raw turn item injection for advanced automation pipelines.
• Windows stability fixes — Rate-limit handling, Guardian timeout resolution, and path-matching fixes for Windows users who've been hitting edge-case crashes since the March rollout.

The marketplace feature carries the most long-term weight. Third-party developers can now build and distribute Codex plugins publicly. A plugin ecosystem on top of an AI coding agent has no direct precedent — the combination of extensibility and autonomous execution is a new category of developer tool.

3 Million Weekly Users and a Growing Ecosystem

As of March 2026, Codex reached 3 million weekly active users, up from 2 million in February — a 50% jump in a single month. Token usage grew 70% month-over-month. Developers using Codex report 55% faster coding velocity on average. Enterprise deployments include Cisco, NVIDIA, Ramp, Rakuten, and Harvey.

These adoption numbers matter beyond marketing. A faster-growing tool generates more community plugins, more Stack Overflow answers, more team-to-team knowledge sharing, and more YouTube tutorials covering edge cases. Claude Code remains an excellent tool — but Codex is growing fast enough that by the end of 2026, ecosystem momentum may matter as much as any benchmark difference.

You can try it right now: Sign into ChatGPT with any account — even the $0 free tier — download Codex Desktop for macOS or Windows, and run one real task: ask it to scan a single file for security issues, or schedule a nightly test run summary. The free tier gives you enough compute to evaluate the full workflow before committing to any paid plan. If you're currently on Claude Code's Max plan at $200/month, this week is the right time to run the comparison side by side.

Related Content — Get Started | Guides | More News