Codex vs Claude Code: API and subscription pricing compared for 2026

If you are judging purely by API cost per token, Codex—especially gpt-5.1-codex-mini—is cheaper than Claude Opus in the official prices compared here. OpenAI lists gpt-5.1-codex-mini at $0.25 input and $2.00 output per 1 million tokens, while Anthropic lists Claude Opus 4.7, 4.6 and 4.5 at $5 input and $25 output per 1 million tokens.

Subscriptions are murkier. Codex usage is governed by five-hour usage windows, and Anthropic has reportedly tested removing Claude Code access from the Pro plan for a small share of new sign-ups.

Quick verdict

• Cheapest API option in this comparison: gpt-5.1-codex-mini, based on OpenAI’s official per-token pricing.
• Codex non-mini is much pricier than Codex mini: OpenAI lists gpt-5.3-codex at $1.75 input, $0.175 cached input and $14.00 output per 1 million tokens on standard API pricing.
• Claude Opus costs more in a simple no-cache example: 1 million input tokens plus 1 million output tokens comes to $30.00 for Claude Opus 4.7/4.6/4.5, versus $2.25 for Codex mini and $15.75 for gpt-5.3-codex.
• For subscriptions, do not compare only the sticker price: OpenAI publishes Codex limits per five-hour window, while Ars Technica reported that Anthropic tested making Claude Code unavailable on Pro for about 2% of new prosumer sign-ups.

API pricing per 1 million tokens

All prices below are in U.S. dollars per 1 million tokens. One important caveat: “cache” does not mean exactly the same thing across providers. OpenAI lists cached input pricing, while Anthropic separates cache writes from cache hits and refreshes.

API/model	Input	Cache	Output	Notes
OpenAI gpt-5.1-codex-mini	$0.25	$0.025 cached input	$2.00	Text-token pricing in OpenAI’s model documentation.
OpenAI gpt-5.3-codex	$1.75	$0.175 cached input	$14.00	Standard pricing on OpenAI’s API pricing page.
Claude Opus 4.7/4.6/4.5	$5.00	$0.50 cache hits and refreshes; cache writes are $6.25 for 5 minutes or $10 for 1 hour	$25.00	Pricing in Anthropic’s Claude API documentation.

For a simple 1 million input tokens + 1 million output tokens workload with no caching, the base cost is:

• gpt-5.1-codex-mini: $2.25.
• gpt-5.3-codex: $15.75.
• Claude Opus 4.7/4.6/4.5: $30.00.

That is not a monthly bill estimate. Real coding-agent spend depends on your input/output mix, how much context the agent rereads, whether caching applies, how large your repository is, and how many edit-test-debug loops you run. Still, for a straight token-price comparison using the official figures above, Codex mini is the lowest-cost option in this set.

Subscriptions: the $20 plan question is not simple

For Codex, OpenAI’s pricing page shows usage limits by tier for local messages, cloud tasks and code reviews per five-hour window. It also says local messages and cloud tasks share a five-hour window, additional weekly limits may apply, Pro $200 receives a usage boost until May 31, 2026, and Enterprise/Edu usage has no fixed rate limits because it scales with credits.

For ChatGPT/Codex plan pricing, one third-party summary says ChatGPT Plus at $20/month includes Codex Web and Codex CLI, while ChatGPT Pro at $200/month includes more usage. Treat that as orientation, not a purchasing guarantee: check OpenAI’s current pricing and your account limits before committing.

For Claude Code, one third-party guide summarizes the subscription paths as Pro at $20/month, Max at $100 or $200/month, or pay-per-token usage through the Anthropic API. But Ars Technica reported that Anthropic tested showing Claude Code as unsupported on the Pro plan for some new sign-ups, while Claude Code remained available on Max plans at $100+/month; Anthropic described it as a small test affecting about 2% of new prosumer sign-ups.

The practical takeaway: do not buy Claude Pro solely on the assumption that Claude Code access is guaranteed. If Claude Code is the reason you are subscribing, verify the current plan details in Anthropic’s pricing page or your account at checkout.

Which is cheaper by use case?

1. Light coding through the API

If your main goal is the lowest token bill, gpt-5.1-codex-mini is the budget pick among the models compared here. Its $0.25 input and $2.00 output pricing per 1 million tokens is far below Claude Opus 4.7/4.6/4.5 at $5 input and $25 output.

2. API work on a non-mini Codex model

gpt-5.3-codex sits in a different price tier from Codex mini. OpenAI lists it at $1.75 input, $0.175 cached input and $14.00 output per 1 million tokens on standard pricing. In the simple 1 million input + 1 million output example, that totals $15.75—well above Codex mini, but still below Claude Opus 4.7/4.6/4.5 in the same no-cache scenario.

3. Coding agents on a monthly plan

With subscriptions, the real question is not just “How many dollars per month?” It is “How many useful agent sessions do I actually get?” Codex has five-hour usage windows and possible additional weekly limits. On the Claude side, third-party summaries point to Pro at $20 and Max at $100/$200, but the Ars Technica report shows that Claude Code access on Pro has been subject to testing for some new users.

If your budget is around $20/month, check two things before choosing: whether the coding-agent feature you need is actually available on that plan, and whether the limits are enough for your normal workflow.

4. Heavy use or team workflows

For heavy coding-agent use, compare ceilings and usage patterns rather than headline prices. OpenAI lists local messages, cloud tasks and code reviews per five-hour window, and notes that weekly limits may also apply. For Enterprise/Edu, OpenAI says there are no fixed rate limits and that usage scales with credits.

For Claude Code, a third-party guide describes Max as a $100/$200-per-month route, and Ars Technica reported that Claude Code remained available on Max $100+/month while the Pro access test was taking place.

How to estimate your own cost

1. Separate input from output. The models compared here price input and output differently, and output is more expensive in all three cases.
2. Model cache costs separately. OpenAI uses cached input pricing, while Anthropic separates cache writes from cache hits and refreshes.
3. Use a real workflow sample. A coding agent that reads many files, writes long diffs and retries after test failures can produce a very different token mix from a simple prompt-response session.
4. Do not treat a cheap subscription as unlimited usage. Codex publishes five-hour usage windows and possible weekly limits, while Claude Code access on Pro has reportedly been tested differently for some new sign-ups.

Bottom line

For the cheapest API route, the answer is clear: gpt-5.1-codex-mini is the lowest-cost option in the official pricing compared here, at $0.25 input and $2.00 output per 1 million tokens.

For monthly subscriptions, there is no universal winner. Codex needs to be evaluated against five-hour usage windows and possible weekly caps, while Claude Code should be checked for current Pro/Max access at the moment you subscribe.

A practical rule of thumb: choose Codex mini if your priority is the lowest API token cost; choose a subscription only after confirming both the current feature access and the usage limits for the way you actually code.