Codex vs Claude Code (2026): pricing, limits, skills and config

Codex and Claude Code are both terminal and IDE coding agents that read a project, run commands and extend themselves with skills, and on the documented facts they differ most in three places: Codex shares one usage allowance with ChatGPT chat, Claude Code bills against a subscription that is separate from Claude chat usage only on Enterprise, and the two read skills and configuration from entirely different paths. This page compares what the vendors publish about plans, limits, skill layout and config files, as of 27 September 2026. It makes no claim about which writes better code, because that needs hands-on testing this page has not done.

Every figure below comes from the vendor's own pricing, docs or changelog, linked in the sources list. Where a number is a documented range rather than a guarantee, it is written as a range.

#What does each one cost?

Both vendors sell the coding agent as part of a general subscription rather than as a separate product. Neither publishes a price for the coding agent on its own.

OpenAI's pricing page lists Free at $0, Go at $8 per month, Plus at $20 per month, and Pro from $100 per month. Business is $20 per user per month on annual billing and $25 billed monthly. Enterprise and Edu are quoted by sales, and the page says those plans have no fixed rate limits, with usage scaling against credits instead.

Anthropic's pricing page lists Free at $0 with Claude Code not included. Pro is $17 per month with the annual discount, billed as $200 up front, or $20 billed monthly. Max starts at $100 per month. Team seats are $20 per seat per month annually and $25 monthly for Standard, and $100 per seat per month annually and $125 monthly for Premium. Enterprise is a seat price plus usage at API rates. Claude Code is included on every paid tier.

The Max plan support article splits Max into two named tiers: Max 5x at $100 per month and Max 20x at $200 per month, described as five times and 20 times the Pro plan's per-session usage allowance. OpenAI's pricing page describes Pro as offering 5x or 20x higher rate limits than Plus and shows separate Pro 5x and Pro 20x columns in its limits table, but this page does not state a price for the 20x tier, because the queue item for this article asked us not to and the only figure we found for it was not on a vendor page.

#How are the usage limits shaped?

Both vendors use a rolling five-hour window with a weekly limit on top, and both publish allowances as estimates rather than counts.

OpenAI's pricing page gives estimated local messages per five-hour period per model and plan. For GPT-6 Luna it lists Plus at 350 to 3,000, Pro 5x at 1,750 to 14,000 and Pro 20x at 7,000 to 56,000. It adds that local messages and cloud chats share your plan's usage allowance, that weekly limits may also apply, and that ChatGPT Work and Codex share usage. Past the included limits, credits pay for further work: the page rates GPT-6 Luna at 2.5 credits for input, 0.25 for cached input and 12.5 for output per million tokens, and notes that GPT-5.6 usage averages 5 to 30 credits per message.

The Max plan article says a session-based usage limit resets every five hours, with a separate weekly limit that resets at a fixed time assigned to each account. It states that Max includes access to Claude Code under one unified subscription, and it does not publish a usage limit specific to Claude Code.

The practical difference is what else competes for the budget. OpenAI documents Codex and ChatGPT drawing on one pool. Anthropic's article documents one per-session allowance covering Claude and Claude Code together, so the same caveat applies; what Anthropic does publish separately is Enterprise, where usage is billed at API rates rather than against a session allowance.

For what a token costs once you are past a subscription, Claude Code's changelog for v2.1.280 on 22 September 2026 records Claude Opus 5.5 becoming the default Opus model at $4 and $20 per million tokens with cache reads at $0.20 per million, and a 1M token context window. The same entry records the default model on Pro and Team Standard changing from Sonnet to Opus.

#Where do skills live in each tool?

This is the sharpest incompatibility between the two, and it is purely a matter of paths.

Codex scans .agents/skills directories at four scopes: $CWD/.agents/skills, $REPO_ROOT/.agents/skills, $HOME/.agents/skills and /etc/codex/skills. Inside a git repository it scans every directory from the working directory up to the repository root. A skill is a folder holding a required SKILL.md, optionally alongside scripts/, references/, assets/ and an agents/openai.yaml. The SKILL.md frontmatter requires both name and description. You trigger one with $skill-name in the Codex CLI or IDE extension, or /skills to browse; in ChatGPT you type @. Codex also selects a skill implicitly when a task matches its description. The initial skills list is budgeted at, in the docs' words, at most 2% of the model's context window, or 8,000 characters when the context window is unknown.

Claude Code reads ~/.claude/skills/<skill-name>/SKILL.md for personal skills and .claude/skills/<skill-name>/SKILL.md for project skills, plus nested <subdir>/.claude/skills/ for skills scoped to a subtree and a managed settings directory for enterprise skills. Frontmatter is optional here: name defaults to the directory name and description is recommended rather than required. You trigger a skill with /name, and Claude can invoke it automatically unless disable-model-invocation is set to true. Claude Code's frontmatter also carries options Codex does not document: user-invocable, allowed-tools, context: fork to run the skill in an isolated subagent context, and arguments for named positional substitution.

So the file format is close enough to share a body, and the discovery paths do not overlap at all. Claude Code's skills page lists only .claude/skills/ and ~/.claude/skills/, and its memory page states that the AGENTS.md loader does not read anything under a .agents/ directory. Neither page documents Claude Code reading .agents/skills/, so a repository that ships Codex skills should not be assumed to expose them to Claude Code. For the wider picture across agents, see where agent skills go by tool and how skills differ from subagents, plugins and MCP servers.

#Which config files does each one read?

Codex uses TOML. User configuration lives in ~/.codex/config.toml, and project-scoped overrides go in .codex/config.toml. Project-local config cannot override machine-local keys: the reference names openai_base_url, chatgpt_base_url, apps_mcp_product_sku, model_provider, model_providers, notify, profile, profiles, experimental_realtime_ws_base_url and otel as ignored when they appear in a project-local file. MCP servers are configured under mcp_servers.<id> sections in the same file. Project instructions come from AGENTS.md, with project_doc_max_bytes capping how much is read from each one and project_doc_fallback_filenames naming alternatives when AGENTS.md is absent.

Claude Code uses JSON, split across five precedence levels. Highest first: managed settings, claude --settings on the command line, .claude/settings.local.json, .claude/settings.json, and ~/.claude/settings.json. MCP servers are stored separately again: local and user scope in ~/.claude.json, project scope in a .mcp.json at the project root that is shared through version control, added with claude mcp add and a --scope flag.

Project instructions are where the two tools nearly meet. Claude Code reads CLAUDE.md from ~/.claude/CLAUDE.md for user scope, ./CLAUDE.md or ./.claude/CLAUDE.md for a project, and a managed policy path per platform, with a documented target of under 200 lines per file. Since v2.1.277 it can also read a repository's AGENTS.md directly, and its documented default is to read AGENTS.md only when no CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md exists in the working directory or above it. A CLAUDE.md that imports AGENTS.md gets both. That means a repository already set up for Codex gets its AGENTS.md read by Claude Code with no extra file, which is the one piece of configuration that genuinely carries across. For the Claude Code side in depth, see the CLAUDE.md memory guide.

#Codex vs Claude Code: the documented differences

What	Codex	Claude Code
Entry paid plan	ChatGPT Go, $8/mo; Plus, $20/mo	Claude Pro, $17/mo annual or $20/mo monthly
Higher individual tier	Pro, from $100/mo, with 5x and 20x rate-limit tiers	Max 5x at $100/mo, Max 20x at $200/mo
Business seat	Business, $20/user/mo annual, $25 monthly	Team Standard $20/seat/mo annual, $25 monthly; Premium $100/$125
Included on the free plan	Reachable on Free; no usage figures published	Not included on Free
Usage window	Estimated messages per five-hour period, weekly limits may apply	Session limit resets every five hours, separate weekly limit
Shares usage with chat	Yes: ChatGPT Work and Codex share usage	One per-session allowance covers Claude and Claude Code
Paying past the allowance	Credits, rated per million tokens	Enterprise usage at API rates
Project skills path	.agents/skills/<name>/SKILL.md	.claude/skills/<name>/SKILL.md
Personal skills path	$HOME/.agents/skills	~/.claude/skills
Skill frontmatter	name and description both required	Frontmatter optional; name defaults to the folder
Skill trigger	$skill-name in CLI and IDE, @ in ChatGPT	/skill-name
Skills list budget	2% of context, or 8,000 characters if unknown	Not documented as a character budget
Config format	TOML	JSON
User config	~/.codex/config.toml	~/.claude/settings.json
Project config	.codex/config.toml	.claude/settings.json and .claude/settings.local.json
MCP config	mcp_servers.<id> in config.toml	~/.claude.json and project .mcp.json
Project instructions	AGENTS.md	CLAUDE.md, and AGENTS.md since v2.1.277

#How well does a migration between them work?

Only one direction is documented as a feature. Codex ships an /import command that pulls a setup across from Claude Code, Claude Cowork or Cursor. Per the docs, settings.json converts to config.toml, MCP server configuration transfers to Codex MCP configuration, project memories become Memories, skills import directly, slash commands convert to skills, hooks transfer to Codex hooks, and chats from the last 30 days import as ChatGPT chats with a cap of 50.

The documented limits matter as much as the list. /import is not available during a running task, in a remote session, or while connected to a local app-server daemon. The Codex CLI also cannot import instruction files or project folders directly: the docs describe those as desktop-app capabilities. Note that skills importing "directly" still lands them where Codex looks, which is not where Claude Code looks, so an imported skill is a copy rather than a shared file.

Anthropic documents no equivalent importer in the other direction. What it documents instead is AGENTS.md compatibility, which covers project instructions but not skills, MCP configuration or hooks.

#Which models does each one run?

Codex's models page lists Astra as its most capable model for complex work, 6 Sol for complex coding and agentic workflows, and 6 Luna as its most efficient model for focused high-volume tasks, with 5.5 retiring on 14 October 2026 and 5.4 and 5.4 Mini still listed. The page says that if you do not specify a model, the ChatGPT desktop app, Codex CLI or IDE extension uses a recommended model, and names Sol Light as the starting preset. It publishes no context window figures. The Codex changelog records GPT-6 Sol and Luna arriving in Codex CLI 0.156.1 on 23 September 2026.

Claude Code's changelog records Claude Opus 5.5 as the default Opus model since v2.1.280 with a 1M token context window. The latest entry at the time of writing is v2.1.283 on 25 September 2026, which added availableModelsMatch and deniedModels managed settings for pinning which model versions a managed deployment may use, and a /doctor prompt-audit command that audits CLAUDE.md files, skills, agents and commands.

One naming caveat: OpenAI's own pages are not consistent about these model names. The pricing page writes GPT-6 Astra and GPT-6 Luna, the models page writes Astra, 6 Sol and 6 Luna, and the changelog writes GPT-6 Sol and Luna. They appear to be the same models under three spellings.

#What this comparison deliberately does not say

Neither vendor publishes a benchmark you can use to compare the two agents on code quality, and a subjective verdict from a page that has not run both on the same repository would be worth nothing. The honest summary of the documented position is narrower and more useful:

• If the deciding factor is a shared allowance, Codex draws on the same pool as ChatGPT chat by OpenAI's own statement, which is the clearest documented constraint on either side.
• If the deciding factor is configuration you already have, AGENTS.md carries across to Claude Code with no work, and nothing else does in either direction without /import.
• If the deciding factor is skills, the two ecosystems do not share a directory, so a team standardising on one path is choosing a tool at the same time.

For a wider comparison that includes IDE-based tools, see Claude Code vs Cursor vs Copilot. The platform hubs for each tool are at Codex and Claude Code.