Codex CLI vs Claude Code: Limits, Cost, Which Is Better

Codex CLI and Claude Code are OpenAI's and Anthropic's flagship terminal coding agents, and neither wins outright. Codex CLI is cheaper per task and rarely hits a usage wall; Claude Code scores higher in independent blind-quality tests but burns through its session limit faster. Anthropic doubled Claude Code's five-hour rate limits in May 2026. Both CLIs have shipped new default models since then, so a comparison written even two months ago is already citing stale versions. This one uses live numbers: Codex CLI on v0.143.0 running GPT-5.6-sol by default, Claude Code at 2.1.177 on this machine as of today.

Key takeaways:

  • Codex CLI's Plus tier costs $20/month with 15-90 messages per 5-hour window; Claude Code's Pro tier is $17-20/month but Anthropic does not publish an exact message or hour cap.
  • In a 36-round blind test, Claude Code won 18 rounds to Codex's 15 (3 ties): a real edge, not the "67%" figure some sites still repeat.
  • Codex CLI uses roughly a quarter to a half the tokens of Claude Code for comparable sessions, which is the main driver of its lower real-world cost.
  • Both support MCP servers; Codex sandboxes at the OS-kernel level (Seatbelt, Landlock, seccomp), Claude Code sandboxes via application-layer permission hooks.

What are Codex CLI and Claude Code

Codex CLI is OpenAI's terminal coding agent: it inspects a local repository, edits files, runs commands, and can operate non-interactively via codex exec for CI pipelines. It defaults to GPT-5.6-sol at medium reasoning effort, and the current release is v0.143.0, installed via a standalone script, npm, or Homebrew (Source: OpenAI Codex docs).

Claude Code is Anthropic's equivalent: an agentic CLI, IDE extension, desktop app, and web surface. It reads a codebase, edits across files, runs shell commands, and integrates with git to open pull requests. It ships as a native installer, a Homebrew cask, or a WinGet package (Source: Claude Code overview).

Operator note (first-hand): running claude --version on this machine returns 2.1.177 (Claude Code). That is newer than the 2.1.165 snapshot independent tester Blake Crosley recorded on June 5, 2026, confirming Claude Code ships meaningful point releases roughly every six weeks. Codex was not installed on this host, so its version comes from OpenAI's published docs instead of a local check. That gap is itself worth knowing before you commit to either tool for a long project.

Codex CLI vs Claude Code: pricing compared

Codex CLI's plans run from free to enterprise, each gated by ChatGPT tier rather than a separate CLI subscription:

PlanPriceCodex CLI access
Free / Go$0 / $8 per monthSame message caps as Plus
Plus$20 per month15-90 GPT-5.6-sol messages per 5-hour window
Pro (5x)From $100 per month75-450 messages per window
Pro (20x)From $200 per month300-1,800 messages per window
Business$20-25 per user/monthTeam-managed, no fixed per-CLI cost

(Source: OpenAI Codex docs)

Claude Code rides on the same subscription as the Claude chat app, so there is no separate line item:

PlanPriceClaude Code access
Pro$17/mo annual, $20/mo monthlyOne weekly cap shared with Claude chat
Max (5x)From $100 per monthTwo weekly caps: all-model and Sonnet-only
Max (20x)From $200 per monthSame structure, higher ceiling
Team$20-100 per seat/monthIncludes Claude Code and Claude Cowork

(Source: Claude pricing)

The structural difference matters more than the sticker price: OpenAI publishes exact message ranges per plan, while Anthropic states only that Claude and Claude Code "share the same usage limits" without naming a number (Source: Claude Code usage limits). If you need to budget spend precisely in advance, that asymmetry favors Codex CLI regardless of which tool writes better code.

Usage limits: which one hits the wall first

Anthropic doubled Claude Code's five-hour session caps for Pro, Max, Team, and seat-based Enterprise plans on May 6, 2026. A week later it added a 50% increase to weekly limits across the same tiers. This is widely reported across independent usage trackers, though Anthropic's own usage-limits page states only that Claude and Claude Code share one pool, without naming exact figures (Source: Claude Code usage limits). Third-party estimates put the Max 5x tier at roughly 140-280 hours of Sonnet usage per week plus 15-35 hours of Opus; Anthropic itself has not confirmed those numbers.

Codex CLI's limits are narrower in scope but easier to plan around: a Plus subscriber gets 15-90 GPT-5.6-sol messages every five hours, a number OpenAI states directly rather than leaving to community trackers (Source: OpenAI Codex docs). One frequently cited Reddit line captures the practical gap: "I've never hit my $20 plan limit. I coded nonstop and never got blocked." (Source: dev.to Reddit sentiment roundup) Claude Code users report the opposite experience on lower tiers, burning 50-70% of a five-hour window on a single complex prompt before the May 2026 doubling took effect.

Model quality: what the blind tests actually show

Several comparison posts still repeat a "Claude Code wins 67% of blind tests" figure sourced from a Reddit sentiment aggregation. The actual methodology tells a different story. Independent tester Blake Crosley ran a 36-round blind test across code review, feature implementation, refactoring, and DevOps tasks: Claude Code won 18 rounds, Codex won 15, and 3 tied, a 50%-to-42% edge, not 67% (Source: Blake Crosley blind-test). Claude Code's advantage concentrated in code review and security verification (8 wins to 4); Codex pulled ahead on DevOps and CI/CD tasks (3 wins to 1).

Token efficiency tells a clearer story. Across a comparable session set, Codex used about 1.5 million tokens against Claude Code's 6.2 million for equivalent work (Source: Blake Crosley blind-test). That gap alone explains most of the cost difference between the two tools, even before per-token pricing enters the math. The community summary of that trade-off holds up: "Claude Code is higher quality but unusable. Codex is slightly lower quality but actually usable." (Source: dev.to Reddit sentiment roundup)

MCP, sandboxing, and extensibility

Both CLIs support the Model Context Protocol (MCP), the open standard for connecting an agent to external tools and data sources, whether that is a local file server, a ticketing system, or a custom API (Source: Claude Code overview; OpenAI Codex docs). Neither tool has an MCP advantage on paper; the difference shows up in sandbox design.

Codex CLI enforces isolation at the operating-system level: Seatbelt on macOS, Landlock or seccomp on Linux, with three explicit modes (read-only, workspace-write, danger-full-access) that a user picks per session. Claude Code takes an application-layer approach instead, gating actions through configurable permission hooks rather than a kernel sandbox (Source: Blake Crosley blind-test). A kernel-level sandbox gives a stronger technical guarantee against a runaway command; an application-layer hook system is easier to customize for a specific team's workflow.

On extensibility, Claude Code has the deeper ecosystem: CLAUDE.md project memory, auto-memory across sessions, Skills for packaging repeatable workflows, more than two dozen hook lifecycle events, and the Claude Agent SDK for building fully custom agents on the same engine (Source: Claude Code overview). Codex CLI's equivalent is an AGENTS.md convention plus MCP server configuration, a lighter but less structured system.

Which should you pick

If your priority is predictable cost and you rarely need deep multi-file refactors, Codex CLI's lower token burn and published message caps make budgeting straightforward. If you need the highest-quality output for security-sensitive review or complex feature work and can tolerate paying for a Max plan, Claude Code's blind-test edge and mature Skills/Hooks ecosystem justify the price.

Blake Crosley, the CTO at Introl who ran the 36-round blind test, runs both tools daily rather than picking one: "I use Claude Code for daily development: feature implementation, code review, multi-file refactors where hooks enforce quality gates at every step. When an external contributor opens a PR, I switch to Codex with --sandbox read-only to review their changes against untrusted code." (Source: Blake Crosley blind-test) That split, Claude Code for trusted first-party work and Codex CLI for reviewing anything you did not write yourself, is a workable default until your own testing says otherwise.

FAQ

Is Codex CLI better than Claude Code?

Neither is better across every dimension. Codex CLI is cheaper, faster, and uses fewer tokens per task, while Claude Code wins more rounds in independent blind quality tests (18 of 36 vs 15). Pick Codex CLI for cost-sensitive or high-volume work and Claude Code when output quality on complex tasks matters more than price.

Are Codex and Claude Code the same?

No. Codex CLI is OpenAI's terminal agent built on GPT-5.6-sol with OS-kernel sandboxing (Seatbelt, Landlock, seccomp). Claude Code is Anthropic's terminal agent built on Sonnet or Opus models with application-layer permission hooks, CLAUDE.md memory, and a broader Skills and subagent ecosystem. They share MCP support but differ in pricing structure, sandbox model, and default reasoning behavior.

Are Codex limits better than Claude Code?

For budgeting, yes: OpenAI publishes exact message ranges per plan (15-90 on Plus, up to 1,800 on Pro 20x), while Anthropic does not disclose exact Claude Code hour or message caps. After Anthropic's May 2026 doubling, Claude Code's real-world ceiling likely rose too, but the lack of published numbers makes Codex CLI's limits easier to plan around in advance.

Can you use Claude in Codex CLI?

Not natively. Codex CLI is built around OpenAI's own models and does not support swapping in Claude as the underlying model. If you want to use both engines from one workflow, run them as separate CLI sessions, or connect Claude Code and Codex CLI to the same MCP servers so both agents see the same external tools and data.

Related coverage

References