TokenSave logo TokenSave

Claude Code vs Codex vs Cursor: Cost and a Real Test (2026)

· October 6, 2026 · Updated October 12, 2026

Claude Code vs Codex vs Cursor: Cost and a Real Test (2026)

Claude Code, OpenAI's Codex and Cursor all sell plans up to $200 a month, so on the pricing page they look interchangeable. They are not. Claude Code and Codex give you a usage allowance that refills on a schedule. Cursor, sold as part of The X Subscription since mid-October 2026, includes a weekly pool for its own Grok and Composer models only and charges Claude, GPT and Gemini separately at API prices. Their free options differ too, and so does how much work each one gets out of a task. This comparison covers the plans, how each tool counts usage, what independent benchmarks show, and what happened when we gave all three the same prompt.

🧮 Open the agent cost calculator →

Claude Code vs Codex vs Cursor at a glance

Claude Code Codex Cursor
Made by Anthropic OpenAI Anysphere (Cursor)
Where it runs Terminal, IDE, Claude app ChatGPT app, CLI, IDE, web Its own code editor
Models Claude only OpenAI only Claude, GPT, Gemini, Grok and Cursor's own
Free option No (Claude Free has no Claude Code) Yes, GPT-6 Luna Yes, Free plan (Cursor's own models)
Entry paid plan Claude Pro $20 ChatGPT Plus $20 Plus $8 (Grok and Composer only)
How usage is counted 5-hour and weekly limits 5-hour and weekly limits on Plus; weekly on Pro Weekly pool for Cursor models; other models at API prices
Shares usage with The Claude app ChatGPT Work and other agent features The Grok apps and Grok Bot

The biggest structural difference is lock-in. Claude Code and Codex each run their own company's models, while Cursor lets you switch between providers in one editor and pays each model's list price.

The plans and prices (2026)

US monthly prices, checked October 11, 2026 (Cursor rechecked October 23 after its plans changed):

Price Claude Code (Anthropic) Codex (OpenAI) Cursor
Free – ChatGPT Free (GPT-6 Luna) Free (limited usage)
$8 – ChatGPT Go (GPT-6 Luna) Plus (Grok and Composer only)
$20 Claude Pro ChatGPT Plus –
$30 – – Premium
$100 Max 5x Pro 100 Super
$200 Max 20x Pro 200 Ultra
$500 – Pro 500 (adds Ultrafast) –

Full details for each tool: Claude Code pricing, Codex pricing and Cursor pricing.

How each one counts usage

Claude Code: 5-hour and weekly limits

You get an allowance per 5-hour session and a weekly cap on top, both shared with the Claude app on web, desktop and mobile. A long chat in the app uses the same allowance as a coding session. Anthropic doubled Claude Code's five-hour limits for Pro and Max on May 6, 2026. Anthropic doesn't publish the allowance in tokens. More: Claude Code usage limits.

Codex: 5-hour and weekly limits, a monthly meter on Free

On Plus and Business, Codex has a rolling 5-hour window, and weekly limits may also apply. Pro plans currently have no 5-hour window, only a weekly allowance. OpenAI publishes ranges per model for Plus: about 5 to 45 local messages per 5 hours on GPT-6 Astra and 350 to 3,000 on GPT-6 Luna. Since April 2, 2026, usage is metered in credits based on API token prices.

On our free account, Codex showed neither a 5-hour nor a weekly meter. It showed one usage bar, labeled one month, with a reset date a month out. More: Codex usage limits.

Cursor: a weekly pool for its own models, API prices for the rest

Since mid-October 2026, each paid Cursor plan includes one weekly pool for Cursor's own models (Grok 4.7, 4.6, 4.5 and Composer 2.5), shared with the Grok apps. Claude, GPT, Gemini and the rest are not included: on Premium and above they are charged at each model's API list price from the first request, paid from prepaid credit. When the pool runs out, you wait for the weekly reset, pay from credit or upgrade. Existing Pro and Pro+ subscribers stay on the older monthly pools.

Same prompt, three tools: our test

Plans and benchmarks only go so far, so we gave each tool the same task between October 9 and 11, 2026. The prompt, word for word:

"Make a brick-breaker game that runs in the browser as a single file named index.html in the current folder. It needs 2 levels, a score display and 3 lives, and it must be playable with the keyboard and the mouse. Write the file and finish."

Same prompt in Claude Code, Codex and Cursor: Sonnet 5.5 medium 37 s, GPT-6 Luna medium 39 s, Grok 4.6 1 min 31 s

Tool · model · effort Plan Time Game code Usage or cost
Claude Code · Sonnet 5.5 · medium Max 5x 37 s 198 lines $0.25 at API prices
Claude Code · Sonnet 5.5 · high Max 5x 58 s 398 lines $0.30 at API prices
Claude Code · Opus 5.5 · medium Max 5x 52 s 358 lines $0.56 at API prices
Claude Code · Opus 5.5 · high Max 5x 1 min 50 s 523 lines $0.75 at API prices
Codex · GPT-6 Luna · medium Free 39 s 398 lines Both Codex runs together: 1% of the monthly free usage
Codex · GPT-6 Luna · extra high Free 39 s 273 lines (included above)
Cursor · Auto (Grok 4.6) Free (Hobby) 1 min 31 s 398 lines 15% of the monthly free usage

Update (October 23, 2026): re-run three times each on Claude Code 2.1.295 with no MCP servers, the same task cost $0.10 (Sonnet 5.5 medium), $0.12 (Sonnet high), $0.26 (Opus 5.5 medium) and $0.35 (Opus high) on average, 53% to 61% less than the October 9 runs in this table. See the effort dataset.

Every run produced a working game that met the prompt: two levels, a score, three lives, and keyboard and mouse controls. The Claude Code costs were measured with ccusage and converted at API prices. The Codex and Cursor figures are what each app's usage meter showed.

What the games were like

The person who ran the tests played every game. In his words, translated from Korean:

The Opus 5.5 runs are reviewed in detail in Claude Opus 5.5 effort levels tested, and the Sonnet runs in Claude Sonnet vs Opus 5.5 for coding.

What the numbers mean (our reading)

What independent benchmarks say

What independent benchmarks say: Metric, Claude Code (Sonnet 5.5, max), Codex (GPT-6.1 Sol, xhigh)

For larger tasks, Artificial Analysis runs both agents on its Coding Agent Index, which combines three coding-agent benchmarks. Its representative configurations are Claude Code with Sonnet 5.5 at max effort and Codex with GPT-6.1 Sol at extra high:

Metric Claude Code (Sonnet 5.5, max) Codex (GPT-6.1 Sol, xhigh)
Coding Agent Index 68 63
DeepSWE v1.1 72% 73%
Terminal-Bench 4.0 66% 55%
SWE-Atlas-QnA 67% 61%
Cost per task $14.19 $1.04
Time per task 1.5 hours 15.5 minutes
Tokens per task 27.7M 3.2M

In our reading, the pattern matches our small test: Codex tends to be lean with tokens, while Claude Code at high effort does more work per task and charges for it. On a subscription you don't see the dollars, but the extra tokens count against your allowance.

Which gives the most for $200?

If you use one agent all day, the flat allowances from Anthropic and OpenAI usually go further than paying API prices, because they aren't tied to per-token rates. Cursor wins on flexibility (any model in one editor) and on seeing the cost of each task.

What a month costs on the API, for reference

From our Claude Code pricing model, 22 working days, with cache reads at 5% of the input price. The range runs from new context billed at the input price to new context billed at the 1-hour cache-write rate, as in our measured runs:

How you use it API with Claude Sonnet 5.5 API with Claude Opus 5.5 Cheapest plan that usually fits
Light: 3 small fixes a day $9–14 $18–28 API, or any $20 plan
Regular: 5 features a day $63–90 $125–179 $20–$100 plan
Heavy: 15 features a day $188–269 $375–537 $100–$200 plan
Agents all day $336–446 $672–893 $200 plan

On Cursor, the API column is roughly what you would pay from prepaid credit, on top of the plan, when using Claude models. On Codex, GPT-6.1 Sol's API prices are the same as Sonnet 5.5's ($2 input and $10 output per million tokens), and GPT-6 Luna's are 95% lower, so a similar workload on Luna would cost far less.

My own setup

I use Claude on the Max 5x plan every day. On Opus 5 I used to run out of my limit; since Opus 5.5 I haven't hit it, though one week recently reached 93%. On the OpenAI side I'm on the free plan, which is where the Codex runs above were done. Cursor was new to me; I installed it for the test above.

Claude Code or Codex: which should you pick?

Verdict

FAQ

Is Claude Code better than Codex? On Artificial Analysis's Coding Agent Index, Claude Code with Sonnet 5.5 at max effort scores 68 against Codex with GPT-6.1 Sol at 63, but at $14.19 per task against $1.04. Which is better depends on whether you value the higher score or the lower cost.

Is Codex cheaper than Claude Code? Both start at $20 a month. Codex also has a free option with GPT-6 Luna, and in independent benchmarks it uses far fewer tokens per task, so the same allowance or budget goes further.

Can I use Claude Code for free? No. Claude Code needs a paid Claude plan (Pro, Max, Team or Enterprise) or an API key.

Is Codex free? Yes, in a limited form. ChatGPT Free includes Codex with GPT-6 Luna in the desktop app. On our account the free plan showed one monthly usage meter.

Can Cursor use Claude and GPT models? Yes. Cursor's paid plans include Claude, GPT, Gemini and Grok models, charged at each model's API price. On the free plan, Cursor picks the model for you.

Do Claude Code and Codex share limits with the chat apps? Yes. Claude Code shares its allowance with the Claude app, and Codex shares one with ChatGPT Work, ChatGPT for Excel and Workspace Agents.

Estimate your own month

Put in your task size and tasks per day in the coding agent cost calculator to see the API cost next to every Claude and ChatGPT plan.

Plan details checked on each company's pricing page on October 11, 2026. Our test is one run per setup between October 9 and 11, 2026; Claude Code costs were measured with ccusage, and Codex and Cursor usage was read from each app's meter.

Sources

More articles

Disclosure: TokenSave gets no funding, credits or sponsorship from AI companies. All measurements use our own paid subscriptions.

Set task size and tasks per day to compare Claude Code's API cost with Pro and Max.

Open the agent cost calculator →