Guide August 19, 2026

Claude Code vs Codex Limits Compared (2026): Which $20 Plan Cuts You Off First

Both cost 20 dollars. Only one of them tells you how much work you actually get before it cuts you off.

Both tools cost $20 a month at the entry paid tier. Both stop you mid-task. The difference is not raw capacity, because neither vendor publishes a number you can compare directly. The difference is how the two limit systems are shaped, and that decides which one fits how you actually work.

This page compares the limit mechanics side by side. For the fixes on each tool, read Stop Hitting Claude Code Usage Limits and Stop Hitting Codex Usage Limits. This one is the choice, not the fix.

Hit Your Codex Token Limit? Do THIS to Reset It

The honest starting point: only one vendor publishes numbers

Anthropic does not publish a message count, a token count, or an hours figure for any plan. The pricing page describes Max as “5x or 20x more usage than Pro” and the support docs describe Max 5x as “five times more usage per session than the Pro plan” (Max plan article). Pro itself has no published baseline, so “5x” multiplies a number you cannot see. The only real reading is the meter inside the product: /usage in Claude Code, or Settings > Usage on claude.ai.

OpenAI does publish per-model message ranges per plan on learn.chatgpt.com/docs/pricing. On Plus, the five-hour window is stated as 10 to 100 messages on GPT-5.6 Sol, 25 to 200 on Terra, and 250 to 2,000 on Luna. The ranges are wide because message cost varies with context size, but a wide published range still beats no range.

That asymmetry is the first real finding. If you need to forecast capacity before you pay, Codex gives you something to forecast with.

Window structure

	Claude Code	Codex
Short window	Rolling 5-hour session window (docs)	Rolling 5-hour window (pricing docs)
Long window	Weekly window, resets at a fixed time assigned to your account, not rolling (Max plan article)	Weekly limits exist, but OpenAI does not publish how the weekly window is anchored
Model-specific cap	Yes. A separate Opus limit sits on top of the shared limits	No separate per-model cap published. Models differ in message cost instead
Published allowance	None	Per-model message ranges per plan

The weekly reset is the sharpest structural difference. Anthropic’s weekly window resets on a fixed schedule tied to your account. The Max plan article states it directly: “Your weekly limit resets at a fixed time each week that is assigned to your account”, and it does not move with your subscription date or your usage. You can learn your reset day and plan around it.

OpenAI states only that “additional weekly limits may apply” on the pricing page and does not publish how that weekly window is anchored. Community posts claim it runs 7 days from your first message of the week; we could not confirm that from an OpenAI page, so treat your reset day on Codex as something to read off /usage, not something to plan a month around.

Both short windows behave the same way in practice: five hours, rolling, consumed by whatever you did five hours ago.

What counts against the limit

The mechanism is identical on both tools and it is the reason most people burn out early. The full conversation is re-sent on every request, and every tool call is another request. A one-line question at hour six of an open session bills for the entire session history.

Anthropic documents this directly. A long session draws usage for the whole conversation at cached rates, and the cache lifetime is one hour on a subscription, dropping to five minutes once you are drawing on usage credits (costs docs). Come back after lunch and your first message reprocesses everything.

Two Claude Code specifics worth knowing before you pick:

Agent teams cost roughly 7x a standard session when teammates run in plan mode, because each teammate holds its own context window (docs). Parallel agents are a real feature and a real multiplier.
/compact is itself a large request because it reads what it summarizes. /clear costs nothing.

On the Codex side the equivalent lever is cached input. Resuming an existing session reuses cache; starting fresh on the same code pays full input rates again.

Shared surfaces

Both vendors pool usage across surfaces, so “just use the web app instead” does not work on either.

Claude Code usage is shared with Claude chat and Cowork on the same plan. Session and weekly limits are shared across all models, so /model does not restore access (errors docs). The exception is the Opus-specific limit, where switching models does keep you working.
Codex draws local CLI messages, IDE usage, and cloud chats from the same five-hour and weekly limits tied to your ChatGPT plan.

One asymmetry: Claude Code’s /usage breakdown is computed from local session history on that machine, so usage from your other devices and from claude.ai is not shown there even though it counts. The meter is narrower than the limit.

What happens when you hit the wall

	Claude Code	Codex
Behaviour	Blocks further requests until the reset time shown in the message	Blocks until the window resets
Reset time shown	Yes, in the message (“resets 3:45pm”, “resets Mon 12:00am”)	Yes, via `/usage`
Model switch escape	Only for the Opus-specific limit	Not a documented escape
Free early reset	No mechanism	Banked resets, when you hold an unexpired one
Paid overflow	Usage credits, via `/usage-credits`	Credits, bought under Settings > Usage
Auto-resume	Desktop app has an “Auto-continue when limits reset” checkbox on the session-limit card, not on the weekly-limit card (errors docs)	Not documented

One Codex-side complication that has no Claude equivalent. OpenAI states that “Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents use a shared allowance and credit pool when those features are available on your plan.” So on the OpenAI side, coding work competes with other agentic features for the same budget. ChatGPT file upload, image, and voice caps are metered separately and do not touch it, and regular Chat Voice does not consume Codex usage. To read where you stand mid-session, /status in the Codex CLI.

Codex is the only one of the two with a free early-reset mechanism. Banked resets are granted, not bought, and they expire 30 days after they are granted. That is a genuine advantage if you work in unpredictable bursts and occasionally need one urgent session finished tonight. The catch is that you cannot rely on holding one.

Claude Code’s counterweight is the auto-continue checkbox in the Desktop app, which retries the interrupted turn after the reset. That is worth more than it sounds for long unattended runs: the work restarts without you sitting there.

Price at each tier

Verified on claude.com/pricing and learn.chatgpt.com/docs/pricing on 2026-08-19.

Tier	Claude	ChatGPT / Codex
Free	$0. Chat only. Claude Code needs a Pro, Max, Team, Enterprise, Console, or cloud-provider account (quickstart)	$0, and Codex is included. OpenAI states Codex “is included across ChatGPT plans, including Free and Go”, with limits varying by plan
Entry paid	No $8 tier	Go, $8/mo
Standard	Pro, $20/mo monthly or $17/mo billed annually	Plus, $20/mo
High	Max 5x, $100/mo. Max 20x, $200/mo	Pro, from $100/mo, sold in 5x and 20x tiers
Team	$25/seat monthly, $20 annually. Premium seat $125 monthly, $100 annually. Anthropic does not publish the Premium seat’s multiplier over a Standard seat	Business, $25/user monthly, $20 annually
Metered escape	Usage credits on top of a plan	API key sign-in, billed at API rates outside plan limits

Two things to notice. OpenAI has an $8 rung that Anthropic does not. And Codex lets you sign in with an API key to bypass plan limits entirely, billing your Platform account at API rates. Claude Code’s equivalent is usage credits stacked on the subscription, or Console authentication for API-billed use.

Anthropic’s own docs put enterprise Claude Code spend at around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users under $30 per active day. That is API-billed usage, not subscription usage, but it is the only concrete per-developer figure either vendor publishes and it is a useful sanity check on whether a $20 plan was ever going to cover full-time agent work.

Which one to pay for

Long sessions, one agent, all day. Claude Code, at Max 5x or higher. The fixed weekly reset lets you plan the week, the Opus-specific limit gives you a fallback lane when the expensive model runs out, and auto-continue recovers unattended runs. At $20 Pro this workload will not fit on either tool.

Short bursts, unpredictable schedule. Codex Plus. Published message ranges let you estimate before you commit, and banked resets cover the occasional night when you need one more hour. The drifting weekly anchor matters less when you are not near the cap.

Parallel agents. Neither $20 plan. Fan-out multiplies context, and Anthropic measures agent teams at roughly 7x a standard session. If you are set on parallelism, budget for Max 20x or move the fan-out onto metered API billing where you can see the cost per run instead of hitting an opaque wall.

Occasional use, cost first. Codex Go at $8. It is the cheapest paid rung in this comparison and Anthropic has no answer to it. One caveat: OpenAI’s per-model Codex message ranges are published for Plus, Pro 5x, Pro 20x and Business, not for Go, so Go is the one tier you cannot forecast before you buy.

You want to forecast before you buy. Codex, for the published ranges alone. Anthropic’s meter is good once you are inside the product; there is nothing to model beforehand.

The uncomfortable summary: at $20 neither plan supports a full day of heavy agent work, and the habits in the two sibling guides move your effective capacity more than the choice between vendors does. Fix the context burn first, then pick the plan.