Claude Code vs Codex: the honest comparison

Two tools dominate agentic coding in 2026: Claude Code from Anthropic, and Codex from OpenAI. Both make the same promise: you describe what you want, the agent writes the code, runs the tests and fixes things until they work.

One disclosure before we compare: at PHTM we build our clients' websites and automations with Claude Code, every day. So we have a bias, and we are naming it. The figures below are verifiable, and Codex is the right choice in several situations we describe honestly.

The short answer

| | Claude Code (Anthropic) | Codex (OpenAI) | | --- | --- | --- | | Entry price | $20/month (no free tier) | Free with a ChatGPT account, or Go at $8/month | | Full range | Pro $20 · Max $100 · Max $200 | Free · Go $8 · Plus $20 · Pro $100 or $200 | | Models | Fable 5, Opus 4.8, Sonnet 5, Haiku 4.5 | GPT-5 "codex" variants, code-optimised | | Where it runs | Terminal, desktop app, web, IDE extensions | CLI, web, IDE, GitHub integration | | Strongest at | Long autonomous work, quality on large jobs | Entry price, parallel cloud tasks |

Pricing in detail

Both vendors bill in dollars, with usage limits per rolling window of a few hours rather than fixed monthly quotas.

Claude Code is included in Claude subscriptions: Pro at $20/month for regular individual use, Max at $100/month for roughly five times the usage, Max at $200/month for roughly twenty times. Teams go through Premium seats at around $100 to 125 per seat. The no-subscription alternative is paying the API per token, which works out cheaper for occasional use and more expensive for heavy use. There is no free tier.

Codex works the other way round: entry is free. The CLI works with a ChatGPT account even on the free plan, the Go tier costs $8/month, Plus $20, and Pro comes in two levels at $100 and $200 since April 2026. If you already pay for ChatGPT, Codex costs nothing extra: it draws on the same subscription. Since April 2026, usage beyond the limits is billed through token credits rather than per message.

On the API side, Claude models range from $1/$5 per million tokens (Haiku 4.5) to $10/$50 (Fable 5), with Sonnet 5 at $3/$15 and Opus 4.8 at $5/$25 in between. Those numbers matter mostly if you wire the tool into your own systems.

The models, in practice

The model choice matters more than the logo on the tool.

On the Anthropic side, Sonnet 5 is the workhorse: fast, affordable, and sufficient for the large majority of coding tasks. Opus 4.8 takes over on long autonomous jobs, the ones where the agent works for an hour unsupervised. Fable 5, the most expensive, is reserved for problems the others cannot crack. Claude Code lets you switch models mid-session, which keeps the cost steerable.

On the OpenAI side, the codex variants of GPT-5 are trained specifically for code and agentic use. They are competitive, particularly on short, well-scoped tasks, and the cloud mode runs several tasks in parallel while you do something else.

On raw quality, both camps publish benchmarks that show themselves winning. Our experience, for what it is worth: on long autonomous sessions, where the agent has to hold an objective across hundreds of steps without drifting, Claude keeps the edge. On a quick, well-described fix, the difference does not justify switching tools.

When to choose which

Choose Codex if you already pay for ChatGPT and want to try agentic coding without spending anything more. If your use is occasional, a few scripts and one-off fixes, the $8 Go tier is the cheapest entry ticket on the market. And if your team lives in GitHub, the code-review integration is native there.

Choose Claude Code if code is your daily production tool. Long autonomous sessions, fine-grained context handling on large repositories, and the extension ecosystem (hooks, skills, MCP servers) make it the most complete tool for intensive professional use. That is what decided our choice: our client automations run on multi-hour jobs, and that is exactly where the gap shows.

The real trap is choosing on benchmarks. Both tools improve every couple of months; today's gap reverses next quarter. Choose on price, on the ecosystem you are already in, and on a one-week trial on your real tasks.

And for an SME without developers?

Honestly: you will probably use neither directly. These tools are for people who build. The question that concerns you is different: does your provider use them well? A website or an automation built with these agents takes noticeably fewer hours than two years ago. If your quotes have not moved, that is a question worth asking.

That is the point of our websites service and our automations: these tools save us hours, and those hours show in our fixed prices. The diagnostic is free: tell us what slows you down.

Frequently asked questions

Can you use Claude Code for free?

No, you need at least the Pro subscription at $20/month or API credits. Codex does have a genuine free tier through a ChatGPT account, with tight usage limits that are enough for a trial.

Do you need to know how to code?

To get real value, yes. The agents write the code, but someone has to check the result, understand what was built and decide what goes to production. An unreviewed agent produces work that looks finished and is not.

What do the usage limits mean in practice?

Both tools cap usage in rolling windows of a few hours, with weekly ceilings. A full-time developer will saturate the $20 tier on either side; that is the signal to move up a level. The $100 tiers cover most intensive professional use.

Can you use both?

Yes, and it is common: nothing makes them exclusive. Some teams keep Codex for small parallel tasks and Claude Code for the big jobs. The cost of trying both for a month is lower than the cost of picking the wrong one for a year.