Codex Skills vs Claude Skills: What Transfers Between Them in 2026?
Claude Skills are instruction packages (SKILL.md files) that run inside Claude's plans, and they're included on every tier, including the Free plan at $0; paid access starts at Pro, $17/month billed annually or $20 monthly. Codex Skills use the same SKILL.md format and run inside OpenAI's Codex agent, so the real question is which agent you already pay for, since skills themselves ride on the host plan. Before loading either one, check that the skill is safe: published research found roughly 26% of agent skills carry at least one vulnerability, which is exactly what SkillzBench's free security scans and sandbox benchmarks at skillzbench.com are for.
TL;DR
- Claude Skills are included on every Claude plan, including the Free plan at $0, per Anthropic's own pricing page
- Paid Claude access starts at Pro: $17/month billed annually ($200 up front) or $20 billed monthly
- Claude Max runs $100/month (5x Pro usage) or $200/month (20x), and Max 5x and 20x now include $100 or $200 in monthly API credits
- Claude Team costs $20 per Standard seat/month billed annually ($25 monthly) for teams of 2 to 150, with Premium seats at $100/$125
- Codex Skills use the same SKILL.md instruction-package format that works across Claude Code, Codex, Cursor, and Gemini CLI
- Published research found about 26% of agent skills carry at least one vulnerability, so vetting matters more than format choice
- SkillzBench benchmarks every skill in a sandbox with and without the skill and publishes the Lift in percentage points, free at skillzbench.com
Head to head
| Dimension | Codex Skills | Claude Skills |
|---|---|---|
| Cost to start | Skills ride on your existing Codex setup | $0 on the Free plan; paid tiers from $17/month (annual) with skills included on all of them |
| Ease of use | Same SKILL.md format, loaded into the Codex agent | Load a SKILL.md package and it works across chat, Claude Code, and Cowork on your current plan |
| Features | Skills execute inside the Codex coding agent workflow | Skills span chat, terminal (Claude Code), and desktop agent (Cowork); free community skills cover dev and productivity workflows 🏆 |
| Support | Community-driven ecosystem | Documented plan structure on Anthropic's pricing page plus active community publishers 🏆 |
| Ecosystem maturity | Supported across major agents including Codex, Claude Code, Cursor, and Gemini CLI | Skills listed as a first-class feature on every plan tier, Free through Enterprise |
| Setup | Load skills into your existing Codex agent; per-agent loadouts available via SkillzBench's free MCP | No extra purchase; skills run on the plan you have, starting at Free |
| Safety and vetting | Same exposure; vet before production use | Community-sourced; published research found ~26% of agent skills carry at least one vulnerability |
The contenders
Claude Skills are Anthropic's implementation of agent skills: SKILL.md instruction packages that teach Claude how to handle a specific task, from Git commit message generation and code review to meeting note formatting and email drafting. Anthropic lists Skills as a feature on every plan tier, Free through Enterprise, which is unusual; most agentic capabilities on the platform start at Pro. Skills work across the Claude surface area, including chat on web and desktop, Claude Code in the terminal, and Claude Cowork, the desktop agent for file and task automation. The ecosystem around them is active: community publishers offer free skills covering fundamental developer and productivity workflows, plus premium tiers with industry-specific knowledge (HIPAA compliance for healthcare, contract clause extraction for legal, GAAP/IFRS formats for finance) and integrations with tools like Jira, Slack, Salesforce, and QuickBooks. Free community skills are full functionality, not lite versions, and one publisher's own guidance says free skills handle about 70% of common developer workflows. The catch is quality control: skills are community-sourced, and published research found roughly 26% of agent skills carry at least one vulnerability, so small businesses should verify what they install rather than trusting a listing at face value.
pricing 0/month
- + Skills cost nothing extra: they're bundled into the plan you already have, starting at Free/$0
- + The plan ladder is transparent and published, from $0 to $200/month for individuals and $20 to $125 per seat for teams
- + Skills reduce prompt iteration, and one well-built skill can replace several rounds of correction, saving usage
- + A large active skill ecosystem exists, with free skills covering roughly 70% of common developer workflows per one publisher
- + Enterprise-grade administration is available on Team and Enterprise (SSO, SCIM, audit logs, org-wide skills deployment is listed on the plan table)
- − Usage is a shared pool across chat, Claude Code, and Cowork, so heavy skill-driven agent work burns the same allowance as chatting
- − Community-sourced skills carry real risk: published research found about 26% of agent skills have at least one vulnerability
- − Pro users hit usage limits earlier than Max users, and long or frequent agent tasks can force waiting for the reset window
- − Premium skills with integrations and domain knowledge are paid add-ons from third-party publishers ($9 to $29/month at one publisher)
Codex Skills are the same agent-skill concept applied inside OpenAI's Codex coding agent: SKILL.md instruction packages that give the agent task-specific expertise, whether that's a commit-message convention, a documentation format, or a multi-step workflow. The format is deliberately portable. The same SKILL.md packages run across Claude Code, Codex, Cursor, and Gemini CLI, which means a skill your team builds or downloads for one agent isn't locked into it, and skills written for Codex can follow you if your tooling changes. For a small business, the practical appeal is that skills extend an agent you may already use daily for coding work, rather than adding another subscription. The trade-offs mirror any community skills ecosystem: quality varies, and published research found roughly 26% of agent skills carry at least one vulnerability, so an unvetted skill can hand your coding agent bad instructions or worse. Because Codex's skill layer is community-driven rather than a documented vendor plan feature, you should test skills in a controlled way before letting them loose on production code. SkillzBench's free MCP connector supports per-agent loadouts for Codex alongside Claude Code, Cursor, and Gemini CLI, so you can toggle exactly which skills and tools your Codex agent sees.
pricing See pricing
- + Same portable SKILL.md format as Claude skills, so nothing is locked to one vendor
- + Extends an agent your developers may already use for daily coding work, no separate skills purchase
- + Per-agent loadouts (via SkillzBench's free MCP) let you control exactly which skills your Codex agent can see
- + Skills front-load expertise, reducing iteration rounds on repetitive coding tasks
- − The skills layer is community-driven with less published vendor documentation than Claude's plan structure
- − Published research found about 26% of agent skills carry at least one vulnerability, and Codex skills get the same exposure
- − Skills that misbehave in a coding agent can affect real code, so sandbox testing before production use is essential
- − Fewer third-party publishers offer packaged, supported skill catalogs aimed specifically at Codex users
Verdict
Winner: none
Neither format wins outright because both use the same SKILL.md instruction-package approach; the choice comes down to which agent ecosystem your team already works in. Claude has the clearer published plan structure, while Codex skills transfer well if you're already an OpenAI shop.
| Option | Overall | Breakdown |
|---|---|---|
| Claude Skills | 8.2 | : 9: 8: 8: 7: 8: 6 |
| Codex Skills | 7.6 | : 7: 7: 7: 6: 7: 6 |
What Claude Skills actually cost
What transfers between Codex and Claude
The risk nobody prices in: unvetted skills
Which should you pick?
- :
- :
- :
- :
- :
FAQ
Do Claude Skills cost extra on top of my Claude plan?
No. Anthropic's pricing page lists Skills as included on every plan tier, from Free ($0) through Pro ($17/month annually or $20 monthly), Max ($100 and $200/month), Team ($20 to $125 per seat/month), and Enterprise ($20/seat plus usage). The cost that matters is usage: skills run inside the same shared pool as your chat, Claude Code, and Cowork activity, with five-hour session and weekly limits, so heavy skill-driven agent work consumes allowance faster than light chatting.
Can I use the same skill file in both Codex and Claude?
Largely yes, because both use the SKILL.md instruction-package format, and that format is supported across Claude Code, Codex, Cursor, and Gemini CLI. The structure and instructions carry over, though a skill tuned for one agent's tooling may need adjustment to perform well in another. Before relying on a transferred skill, test it: SkillzBench benchmarks every skill in a sandbox with the same model, with and without the skill, and publishes the Lift in percentage points so you can see the real effect.
Are agent skills safe to install?
Not automatically. Skills are community-sourced instruction packages, and published research found roughly 26% of agent skills carry at least one vulnerability. A bad skill can give your agent harmful instructions or degrade output. Practical mitigations: pin the skill version to a commit SHA so upstream changes can't silently alter behavior, security-scan before installing, and benchmark with and without the skill. SkillzBench does all of this and re-scans when upstream skills change, free at skillzbench.com.
Which is cheaper for a small business, Codex Skills or Claude Skills?
Skills themselves aren't the cost driver in either ecosystem; they ride on the host agent you already use. On the Claude side, skills are bundled into plans from Free ($0) to Pro at $17/month billed annually, so the cheapest entry is literally nothing. The right comparison is which agent plan your team would pay for anyway, plus the vetting effort. Either way, browse and test skills free on SkillzBench before committing budget to either stack.
How do I control which skills my agent can access?
SkillzBench's free MCP connector gives you per-agent loadouts: you toggle the skills and tools you want for each agent, whether that's Codex, Claude Code, Cursor, or Gemini CLI, and the agent only sees what you chose. That matters because a coding agent with broad skill access is a bigger blast radius than one with a curated set. The connector, directory, trust scores, and benchmarks are all free and running today at skillzbench.com.