Claude Code and Codex both start at $20 a month, both top out around $200, and both will happily burn through an API budget if you let them. The sticker prices are nearly identical. What you actually pay is not, because the two agents spend tokens very differently. This guide compares every published price as of late September 2026, separates what the vendors state from what third parties measured, and then does the part most comparisons skip: cost per finished task.

What do Claude Code and Codex cost per month?

Both tools ship inside a chat subscription you may already pay for. Claude Code comes with Claude Pro and above; Codex comes with every ChatGPT plan, including the free one. All figures below are published on Anthropic's pricing page and OpenAI's plan page.

Two differences stand out. Codex has a real free and an $8 on-ramp, so a developer can try it without a card; Claude Code starts at $20. And Anthropic's Enterprise tier is the only one here that separates the seat fee from usage, billing the agent's work at API rates on top.

What do you actually get for the subscription?

This is where the two vendors diverge in a way that matters for planning. Both meter usage in rolling five-hour windows with a weekly cap on top. Only one tells you the numbers.

OpenAI publishes ranges per five-hour window, per model. On Plus: 5 to 45 messages on GPT-6 Astra, 15 to 150 on GPT-6 Sol, and 350 to 3,000 on GPT-6 Luna. The Pro tiers multiply that by 5x or 20x. The ranges are wide because a message that reads a large repository costs far more than a one-line fix, but they are at least ranges you can plan around.

Anthropic describes Claude Code allowance only relatively: Pro gets "at least 5x more usage per 5-hour session than Free", Max 5x and Max 20x get five and twenty times Pro. Anthropic raised the five-hour limits on its paid plans during 2026, but it publishes no absolute prompt or token count for them or for the weekly cap. Community estimates put Pro around 90 prompts per window, which is folklore rather than a documented figure.

The practical upshot: on Codex you can see which model is eating your window and switch to a cheaper one. On Claude Code you find the ceiling by hitting it.

What does each model cost on the API?

Subscriptions are flat. The moment you run either agent headless, in CI, or at team scale on a pay-as-you-go key, you pay per token. Published list prices, per million tokens:

Sources: Claude pricing, OpenAI API pricing. OpenAI's GPT-6 prices are for standard context; above its long-context threshold they rise, for GPT-6 Sol to $4 in and $15 out.

September 22 reshaped this table. Anthropic cut its default Opus tier by 20% per token (Opus 5.5 at $4/$20, with cache reads down to $0.20), and OpenAI halved the price of Sol and Luna and called the new rates permanent. At the frontier the two are now priced identically. At the default tier Codex is half the price per token. At the bottom Codex has no Anthropic equivalent: GPT-6 Luna costs a tenth of Haiku 4.5.

Why is the cheaper token not the cheaper task?

Because price per token and tokens per task move independently, and the second one varies more than the first.

Independent measurements by Artificial Analysis put the output per agentic coding task, at maximum effort, at roughly 119,000 tokens for Opus 5.5, 78,000 for Fable 5.1 and 27,000 for GPT-6 Astra. Claude models think out loud, re-read files and verify their work; Codex models tend to act and stop. That behavior shows up directly on the bill. An earlier head-to-head by Leanware, on the previous model generation, found the same shape: building the same Figma-style clone took about 6.2 million tokens in Claude Code versus 1.5 million in Codex.

Cost per task on Terminal-Bench 4.0, from the same independent runs:

Read that table in both directions. GPT-6 Sol does a task for a fifth of the price of Opus 5.5 at medium effort. Opus 5.5 at medium also solves a third more tasks, and a failed task costs you its tokens plus your time to redo it. The honest metric is cost per solved task, and at that level the gap narrows sharply for hard work while staying wide for routine work.

A rough budgeting rule falls out of it. At those per-task costs, $20 of API spend buys about 15 medium-effort Opus 5.5 tasks or about 80 GPT-6 Sol tasks. A subscription that covers more than that per month is the better deal; a workload that runs in CI all day is not a subscription workload at all.

Which one does better work?

It depends on the benchmark, and on who ran it. Vendor-reported Terminal-Bench 4.0 scores have Opus 5.5 at 66.4% against 57.9% for GPT-6 Astra and 55.8% for Fable 5.1. On DeepSWE, a harder software-engineering suite, Opus 5.5 and GPT-6 Astra are effectively tied at 74.2% and 74.1%, with GPT-6 Sol at 68.8%. Treat the vendor-run numbers with care: each comes from the vendor's own harness.

The pattern that survives every source: Claude's models keep gaining as you give them more effort on a hard problem, and Codex's cheaper models are good enough for a large share of everyday tasks at a fraction of the cost. That is less a verdict than a description of two different price-performance curves.

How do they differ as tools?

Price aside, they are built for different ways of working.

Neither column is the winner, which is why this grid has no accent. Many teams already run both: Claude Code for the change that needs judgment, Codex for the ten that need throughput.

Which should you pick?

If you only take one thing away: compare cost per solved task on your own repository, not price per token on a pricing page. Both vendors make that easy to measure, and the answer is different for every codebase.

What does this mean if you build agents?

The same economics apply to every agent you ship, not just the one writing your code. Tokens spent reading and re-reading a page, retrying a failed action or guessing at what happened are the most expensive tokens there are, and they buy nothing. An agent that gets typed responses, stable error codes and events instead of screenshots spends its budget on the task. That is the argument behind why AI agents need an API, not a browser, and it is how Curviate exposes LinkedIn to both of these tools: the same actions through the CLI a coding agent already knows how to call, or through MCP for agents that speak it. See every interface for the full picture.

FAQ

Is Claude Code more expensive than Codex?

At the same subscription tier, no: both cost $20, $100 and $200 a month. On the API, Claude's default model costs twice as much per token as Codex's, and Claude models typically use more tokens per task, so heavy API workloads usually cost more on Claude Code. For hard tasks, Claude's higher solve rate narrows that gap.

Is Codex better than Claude Code?

It depends on the benchmark and who ran it. On vendor-reported Terminal-Bench 4.0 scores, Opus 5.5 leads at 66.4% against 57.9% for GPT-6 Astra and 55.8% for Fable 5.1; on DeepSWE, Opus 5.5 and GPT-6 Astra are effectively tied around 74%. The pattern that holds across every source: Claude's models keep gaining as you give them more effort on a hard problem, while Codex's cheaper models are good enough for a large share of everyday tasks at a fraction of the cost. Neither wins outright.

Is Codex free?

Codex is included in the free ChatGPT plan with low usage limits, and in Go at $8 a month. Claude Code has no free tier; it starts with Claude Pro at $20 a month.

What are the Claude Code usage limits?

Anthropic meters Claude Code in five-hour windows plus a weekly cap and describes them relatively: Pro gets at least five times the free plan, Max 5x and Max 20x get five and twenty times Pro. It publishes no absolute number.

Can I use Claude Code and Codex together?

Yes, and many teams do. They read different instruction files (CLAUDE.md and AGENTS.md), so a repository can carry both. A common split is Claude Code for interactive, judgment-heavy changes and Codex's cloud agent for parallel, well-specified tasks.