Neither Anthropic nor OpenAI publishes a fixed API-dollar allowance for its individual coding subscriptions. To get a useful comparison, run npx ccusage@latest claude monthly and npx ccusage@latest codex monthly, take each Cost (USD) total as M, and screen that number against the $20, $100, and $200 seats. Stay on the cheaper seat until it actually blocks paid work. M can exceed the seat price without becoming a bill.
Prices, commands, and source quotations below were checked on 29 August 2026. Confirm the plan shown to your account before buying.
Why feel cannot pick a seat
Claude and ChatGPT sell relative usage, not a dollar wallet.
Anthropic’s pricing page lists Pro at $20 per month ($17 per month on annual billing, $200 billed up front) and Max from $100 per month, with a choice of 5x or 20x more usage than Pro. Its FAQ says usage depends on conversation length, model, and features, so there is no fixed message count.
OpenAI’s Codex pricing page lists ChatGPT Plus at $20 per month, Pro 5x at $100 per month, and a Pro 20x $200 per month tier. Its plan comparison describes the Pro tiers by relative usage, not an included API-dollar balance.
Neither company converts 5x or 20x into “this seat includes $X of API tokens.” That is why a local API-equivalent estimate is useful as a comparison input, but not as a subscription balance.
The five-hour window and weekly cap are a separate decision: see ChatGPT’s five-hour Codex window and Claude Pro limits, Max, and usage credits. This guide only converts local usage into API-equivalent dollars and maps that number onto current individual seats.
What ccusage actually measures
ccusage is a local CLI. Its documented problem list includes estimating what the same work would cost on a pay-per-use plan. It reads usage files already on the machine and estimates cost from token usage and model pricing data. It does not need access to an Anthropic or OpenAI billing account.
Run it without a global install. The getting-started page lists npx ccusage@latest alongside bunx ccusage and pnpm dlx ccusage. For a seat decision, namespace the two coding CLIs instead of mixing every detected agent:
npx ccusage@latest claude daily
npx ccusage@latest claude monthly
npx ccusage@latest claude monthly --breakdown
npx ccusage@latest codex daily
npx ccusage@latest codex monthly
npx ccusage@latest codex monthly --breakdownDo not confuse this npm CLI with the different PyPI project also named ccusage, whose own description identifies it as a Claude Code rate-limit monitor.
The ccusage introduction documents three important limits:
- Local files only — another laptop does not appear.
- Language-model tokens — tool calls such as Web Search are not included.
- Estimates — costs may not reflect actual billing.
Claude Code transcripts are not kept forever. Anthropic’s settings reference sets cleanupPeriodDays to a default of 30. A monthly Claude report on an unmodified machine may therefore cover roughly the last 30 local days rather than a full invoice month. Raise that setting before you need a longer history.
Codex support is marked experimental on the Codex data-source page. OpenAI’s Codex CLI started emitting token_count events in a 6 September 2025 commit; earlier sessions have nothing to price. The same page says Fast pricing uses a model-specific multiplier only when one is published. GPT-5.6 Sol, Terra, and Luna use the documented 2× API Priority rate. That is distinct from the 2.5× ChatGPT credit-consumption rate shown for GPT-5.6 Fast mode: ccusage’s costUSD remains an API-equivalent estimate, not a ChatGPT credit balance.
Hacker News posted the tool on 18 July 2025 as item 44610925, which currently shows 75 points and 30 comments. Author ryoppippi confirmed authorship in the thread. The discussion includes an early self-report comparing ccusage-style usage with a $100 Max seat.
Read three numbers, not the whole table
The getting-started report labels the money column Cost (USD) and defines it as estimated cost based on model pricing. The monthly-report page tells readers to use the total to understand what usage-based pricing would cost. It also labels that comparison Subscription Value.
For each CLI, write down:
| Label | Where | What it is for |
|---|---|---|
| M | monthly Total / this month’s Cost (USD) | The API-list equivalent of the month you actually ran |
| Busy-day cost | daily peak Cost (USD) | Whether one sprint day, rather than the average, is what will hit a window |
| Models | monthly --breakdown | Whether M comes mostly from expensive or cheaper models |
Keep two M numbers. Claude Code and Codex are separate seats. Add them only if you intend to pay for both subscriptions.
ccusage monthly --last 1 shows this calendar month so far. If Claude logs have already aged out, M is a floor rather than a complete month.
Current individual seats in USD
Checked 29 August 2026. Claude prices exclude applicable tax, and Claude Max mobile pricing can differ. Confirm the price shown to your account before buying.
| Seat | Official USD | Usage claim, not a dollar bucket | Source |
|---|---|---|---|
| Claude Pro | $20/mo; annual $200 up front ($17/mo) | More usage than Free; Claude Code included | Claude pricing |
| Claude Max 5x | $100/mo on web | Five times more usage per session than Pro | Max plan help |
| Claude Max 20x | $200/mo on web | 20 times more usage per session than Pro | Same Max help page |
| ChatGPT Plus | $20/mo | Codex included; local-message ranges use a five-hour window | Codex pricing |
| ChatGPT Pro 5x | $100/mo | 5x tier; higher Codex ranges than Plus | Same Codex pricing page |
| ChatGPT Pro 20x | $200/mo | 20x tier; higher Codex ranges than Pro 5x | Same Codex pricing page |
The Max help page gives the web prices as $100 per month for Max 5x and $200 per month for Max 20x, and warns that mobile pricing can vary by app platform.
On the writer’s 28 August fetch, chatgpt.com/pricing was localized to PLN rather than USD: Plus showed PLN 99.99 per month and Pro started at PLN 439 per month. Those figures are not dollar prices. The USD comparison above uses OpenAI’s Help and Codex pricing pages instead.
The upgrade gaps before tax are:
- $20 to $100: $80
- $100 to $200: $100
Neither company publishes an included-API-dollar cap for these individual seats. Do not invent one from 5x or 20x.
List rates explain why M can look large. Anthropic’s API pricing lists Fable 5 at $10 input / $50 output, Opus 5 at $5 / $25, Sonnet 5 at $2 / $10, and Haiku 4.5 at $1 / $5 per million tokens. OpenAI API pricing lists short-context GPT-5.6 Sol at $4 / $20, GPT-5.6 Terra at $2 / $12, and GPT-5.6 Luna at $0.20 / $1.20 per million input/output tokens. OpenAI says Sol’s promotional pricing is available at least through 21 November 2026. These are pay-per-use prices, not subscription entitlements.
Decision rule: M screens the seat; a block buys it
Default to the cheaper seat until it stops paid work, even if M already exceeds $20 or $100.
That is the opposite of “M is $600, so I owe $600.” Subscription usage is sold in relative buckets. API-list M is a counterfactual. The monthly-report page describes what the same usage would cost under usage-based billing, not what the seat contains.
Use M only as a first screen. Then check whether you actually hit a limit.
| Monthly API-equivalent M | First screen | Buy the next seat only if |
|---|---|---|
| M well below $20 | Stay on $20 (Claude Pro or ChatGPT Plus) | The $20 seat still stops a deadline |
| $20 < M ≤ $100 | Still start at $20 | Idle or missed work from limits costs more than the $80 gap |
| $100 < M ≤ $200 | Screen $100 (Max 5x or Pro 5x) | $100 still blocks; the next gap is $100 |
| M > $200 | Screen $200 (Max 20x or Pro 20x) | $200 still blocks; then compare usage credits or API prepaid credit rather than inventing a higher individual seat |
Work the two CLIs separately. A $180 Claude M does not justify ChatGPT Pro, and a $180 Codex M does not justify Claude Max.
Busy-day cost is the tie-break when the monthly average looks cheap. A $12 average day and a $90 sprint day are different workload shapes: the sprint day is more likely to fill a five-hour window.
--breakdown matters because equal token counts can have very different list-price estimates. A Haiku- or Luna-heavy month can print a modest M; an Opus-, Fable-, or Sol-heavy month can print a large M from fewer sessions. Do not upgrade to “buy” a model that the cheaper seat already includes if that seat is not hitting a cap.
If you never hit a limit, keep the cheaper plan at the next billing boundary. M is then evidence that the seat cost less than API list for that workload, not evidence that you underpaid.
Calibration, not quotas
Three public data points show why M can exceed the seat price. None is an official included-dollar figure. Do not copy them onto a pricing page.
-
On the Hacker News ccusage thread, user extr reported running $600–$800 per month in API-equivalent usage on the $100 Max plan while not pushing it to the limit. That is one self-reported account from July 2025.
-
SemiAnalysis wrote on 10 June 2026 in this post that it bought each Anthropic and OpenAI subscription tier, ran long-horizon coding tasks until weekly limits were exhausted, and found the subscriptions more generous than the common belief that a $200 plan tops out around $2,000 per month at API prices. Treat that as an independent third-party measurement. The attached chart was not OCR’d, so this article takes no extra multiples from the image. It is not an Anthropic or OpenAI quota.
-
Cursor forum thread Cursor $200 vs. Claude Max, post 17, says: “My usage on Cursor was $1500-2000/month. The same usage on Claude Code is $200/month.” The poster also reported never hitting Claude Code’s rate limits. This compares one person’s Cursor API-style spend with a $200 Claude Code seat. It does not mean Max 20x contains $1,500–$2,000 of Anthropic API usage.
The safe reading is that API-list dollars are a poor predictor of which seat you need, while still showing why a flat seat can beat pay-per-use for some workloads.
Two traps
Trap 1 — subscription M is not an API invoice.
ccusage estimates cost from tokens and model pricing data. Claude Code and Codex subscriptions do not settle the reported M as an invoice. The HN $600–$800-on-$100 report, the SemiAnalysis test, and the Cursor $1,500–$2,000-versus-$200 report point in the same direction, but remain third-party calibration rather than official allowances. Limits and subsidies can change. Recheck M and the live Usage page after a pricing or limit change, and do not assume today’s relationship survives the next model launch.
Trap 2 — M does not show the window or weekly cap.
Anthropic says session usage resets every five hours, with weekly limits layered on paid plans. OpenAI’s Codex pricing table gives GPT-5.6 Sol local-message ranges per five-hour window: Plus 10–100, Pro 5x 50–500, and Pro 20x 200–2,000, with additional weekly limits possible. A low monthly M can still stop a one-day sprint; a high monthly M can still fit if work is spread across resets. For the mechanics, see the Codex five-hour decision and Claude Pro’s $20 wall.
Checklist
- Run the two
npx ccusage@latestmonthly commands on the machine that does the work. - Record M_claude and M_codex separately. Do not add them unless you pay for both seats.
- Screen M against $20 / $100 / $200. Do not treat M above the seat price as an automatic upgrade order.
- Upgrade only when the cheaper seat blocks paid work by more than the $80 or $100 gap.
- Recheck after one billing cycle using a new monthly report, not the HN or SemiAnalysis figures.
- If $200 still blocks you, compare credits or API prepaid credit against measured overflow, and remember that API prepaid credit expires in one year.
Sources checked 29 August 2026: ccusage introduction, getting started, monthly reports, Codex data source, Claude Code settings, Hacker News item 44610925, SemiAnalysis post, Cursor thread 148298, Claude pricing, Max plan help, Codex pricing, OpenAI API pricing, Anthropic API pricing, and PyPI ccusage.