Claude Max 20x isn't 20x. We checked 60 days of logs against Codex

Updated September 27, 2026 · first published September 27, 2026

Claude Max 20x and ChatGPT Pro 20x (Codex) both cost $200 a month, and both are worth many times that at API prices. Claude Max 20x holds more raw API value: we measured its weekly cap at about $1.86k of API-equivalent usage, against a projected $1.34k for ChatGPT Pro 20x. Codex finishes more work with that money, because GPT models needed 3.3 to 4.2 times fewer dollars for the same coding task in independent tests. So at $200, ChatGPT Pro 20x buys about 2.4 times more finished tasks, as long as OpenAI's 20x multiplier holds for the weekly window as well as the 5-hour one. At $100 the two plans are close to even. Claude Max 5x has the edge in our data, because its weekly cap never bound.

The rest of this page explains how we got those numbers and how to check them against your own logs. It is built on 60 days of one heavy user's Claude Code and Codex logs, priced token by token at the API list price in force on each day, and set against every public measurement we could find. It is meant to settle the decision for someone who genuinely cannot choose between the two plans. The numbers below are the reason for each call.

If you only came for one answer, here they are in the words people search with:

• Is Claude Max 20x really 20x? Only per 5-hour session. Per week it gave at most about 1.5x a Max 5x week in our logs, and about 2x in independent tests.
• Which $200 plan gets more work done? ChatGPT Pro 20x with Codex, by about 2.4x, if its weekly limit really scales 20x. It still wins down to about 7x.
• Which $100 plan? Claude Max 5x, narrowly. Its weekly cap never bound.
• Is Codex cheaper than Claude Code per task? Yes, by 3.3x to 4.2x in two independent tests.
• Should you split $200 across both vendors? It beats Claude Max 20x on finished tasks, but not ChatGPT Pro 20x. Details below.

Why can’t you compare the two “20x” plans directly?

Both vendors sell a $200 plan labelled “20x”, and a $100 plan labelled “5x”. The labels invite a comparison that does not hold, because the two companies are multiplying different things.

• Anthropic sells Claude Max 5x ($100) and Max 20x ($200). The multiplier applies to the 5-hour session relative to the $20 Pro plan. A separate weekly cap sits on top of it, and Anthropic does not publish its size. A heavy user of Claude Code runs into the weekly cap, so the headline multiplier says little about what that user gets.
• OpenAI sells ChatGPT Plus ($20), Pro 5x ($100) and Pro 20x ($200). Its Codex usage page lists per-window message ranges that scale by exactly 5x and 20x between tiers. It also prices extra Codex credits at API parity. OpenAI does not state whether the weekly window scales by the same factor as the 5-hour window.

So the useful question is not “which 20x is bigger?” It has two parts:

1. How many dollars of API-equivalent compute does a full week of each plan hold?
2. How many finished tasks does that money buy, given that the two model families spend very differently to reach the same result?

The first part can be measured from local logs. The second needs independent benchmarks. We did both.

How do you measure a subscription limit in dollars?

Both command-line agents keep complete local logs. Claude Code writes every assistant turn to ~/.claude/projects/**/*.jsonl, with a usage block that splits fresh input, output, cache reads, 5-minute cache writes and 1-hour cache writes. Codex writes ~/.codex/sessions/**/*.jsonl with cumulative token counts per session. Each Codex event also carries a rate_limits object: the percentage used and the reset time for both the 5-hour and the weekly window.

For every turn, we priced the tokens at that model's API list price:

cost = input·pin + output·pout + cache_read·pcr + write5m·pw5 + write1h·pw1

Turns were de-duplicated on message ID and request ID, because Claude Code logs streaming chunks more than once. Codex reports cumulative totals, so we priced the difference between consecutive events and started a new count whenever a session reset its counter. Reasoning tokens are already inside Codex's output count, so they are not added twice.

A plan's limit can then be expressed in dollars in two ways.

• Claude. When a limit binds, Claude Code prints “You've hit your session limit” or “…weekly limit”, followed by the reset time, and that message lands in the log. Weekly windows reset on Friday at 13:00 UTC. Summing API dollars from the start of the window to the logged hit gives the value of the cap at that moment.
• Codex. Codex logs its gauge on every turn, so no hit is needed. For each weekly window, value of a full window = 100 × (dollars spent in the window ÷ percentage points consumed). We kept only windows where at least 30 points were used, so rounding noise on small moves cannot inflate the result.

How much usage is in a Claude Max 20x week?

About $1.9k of API-priced usage at today’s limits, $1.74k to $2.2k depending on the model mix. Our user hit the weekly cap twice in August, at $2,085 and $2,379, during a promotion that ran limits at 150%.

Our user hit the Max 20x weekly cap twice in August: at $2,085 on 14 August and at $2,379 on 19 August. The week that followed stopped at $2,091 without a hit, and the week of 28 August went to $2,636 without a hit either. The cap is therefore not a fixed dollar amount. Anthropic meters the plan in its own units, and those units map onto list prices differently for each model. A week weighted towards Fable, which costs twice as much as Opus per token, reaches the cap on fewer dollars than a week that is almost all Opus. That fits the two hit weeks: the earlier hit week had almost twice as much Fable spend as the later one ($628 against $326) and hit the cap on $300 less.

All four of those weeks fell inside a promotion that raised limits to 150% of normal. Limits have since returned to 125%. Scaling the observed range by 125/150 puts today's Max 20x weekly cap at roughly $1.74k to $2.2k of API-equivalent usage. We use $1.86k, the scaled mean of the two hits, as the central estimate.

Across whole months the picture is the same. In August, on Max 20x, our user used $8,676 of API-equivalent Claude usage for a $200 fee, which is 43 times the price. From 1 to 27 September, mostly on Max 5x at $100, the figure was $4,511, or 45 times the price. Finout's survey of Claude Code costs puts heavy users at $600 to $1,500 a month of API value, so this user sits well above typical usage. Treat the numbers here as the ceiling of what a plan can deliver.

Why is a Claude Code bill mostly cache, not output?

Because the agent re-reads its whole working context on every step. Cache reads were 96% of the tokens our user consumed in August, and output tokens were only $1,038 of a $6,986 Opus bill.

An agentic coding session re-reads its whole context on every step: system prompt, tool definitions, the open files and the conversation so far. Almost none of that is new input. It comes back as a cache read at a tenth of the input price, but at 7 billion tokens a month that tenth adds up to half the bill. Claude Code also writes its cache with a one-hour lifetime, which costs twice the input price, so a session that keeps extending its context pays that premium again and again.

For FinOps this matters directly. A forecast built on output tokens undercounts agentic coding by a factor of about seven. Output was 0.6% of the token volume and 15% of the cost. Any showback or chargeback model for coding agents has to price cache reads and cache writes as separate line items, or it will be wrong by an order of magnitude.

It is also why a Claude subscription looks so generous next to the API. Anthropic meters the subscription in its own units. At list prices, cache-heavy traffic is exactly the traffic that turns a $200 fee into $2,000 a week of API-equivalent value.

Which limit does Claude Max 5x hit first: session or weekly?

The 5-hour session limit. It stopped our user twelve times in September, after a median $37 of work, while the Max 5x weekly cap was never reached.

On Max 5x, the weekly cap never bound, not even in the $1,461 week. The 5-hour session limit bound twelve times in September. The dollars used in the window before each hit ranged from $5 to $127, with a median of $37 and a mean of $44. The spread is wide because a session can begin while the user is already partway through the previous window. The two largest values, $93 and $127, are the best guide to what one full and deliberate Max 5x session holds.

That produces a different pattern of use. Max 20x users run into a weekly limit, while Max 5x users run into a limit every afternoon. If your work comes in long, uninterrupted sessions, the 5-hour limit on Max 5x will cost you more than its weekly cap does.

How much Codex usage do ChatGPT Plus and Pro include per week?

A full Plus week of Codex was worth a median $67 at API prices since August 2026. Scaled by OpenAI’s published multipliers, that projects to about $335 a week on Pro 5x and $1,340 on Pro 20x.

We measured 27 weekly Codex windows on the Plus plan. Pooled across all of them, one full weekly window was worth $103 at API prices, and one full 5-hour window was worth $29.8. Pooling hides a clear break, though. From March to July the median window held $115. Since August it has held $67, and the recent windows are much less scattered: seven in a row fell between $61 and $81.

This drop came mostly from prices, not from a smaller quota. In August the default Codex model moved to Sol, which lists at $2 input and $10 output, against $5 and $30 for GPT-5.5. If OpenAI's weekly quota tracks compute, cheaper tokens mean fewer dollars per window. For forecasting, use the recent figure. We take $67 as the value of one Plus week.

We have not run Pro 5x or Pro 20x. The Pro figures below are projections, calculated as $67 times OpenAI's published multiplier: $335 a week for Pro 5x and $1,340 a week for Pro 20x. The whole comparison depends on that assumption, and we test it below.

Is Claude Max 20x worth it over Max 5x?

For most heavy users, no. Max 20x costs twice as much, but its weekly allowance is only about 1.5x to 2x a Max 5x week, not the 4x the labels suggest.

If the Claude tiers scaled by their labels, a Max 20x week would hold four times a Max 5x week. The evidence points to about half that at best.

• Codeongrass tested both Claude tiers and found the Max 20x weekly allowance to be about 2x Max 5x.
• In our logs, today's Max 20x cap (about $1.86k at 125%) sits at most 1.5x above the $1,217 that a Max 5x week delivered (the $1,461 promo week, scaled to 125%) without hitting its cap. It is an upper bound because the 5x cap was never reached. The real ratio could be lower still.

The practical conclusion is the one that matters most in this study. On Claude, Max 5x is the better buy per dollar, because it gives you half the price and well over half the weekly capacity. The extra $100 for Max 20x mainly buys longer 5-hour sessions. That is worth paying for only if the session limit, not the weekly cap, is what stops you.

Is Codex cheaper than Claude Code for the same task?

Yes, by 3.3x to 4.2x in two independent tests, because GPT-6 Sol reaches the same result with far fewer tokens. Claude still leads on benchmark quality.

API-equivalent dollars measure what the vendor gives up, not what the user gets. A model that needs four times as many tokens to finish the same job turns a larger allowance into the same amount of output. Two independent sources measure that directly.

Quality does not make up the gap in either direction. On Terminal-Bench 4.0, Opus 5.5 scores 66.4% and GPT-6 Sol 60.5%. That is a real lead, but a six-point lead does not offset a 3.3x cost difference for routine work. Claude's lead is worth paying for on hard, ambiguous, multi-step problems. It is not worth paying for when the job is clearing a queue of well-specified tickets.

Take the DataCamp figure as the unit of work. A Tetris-sized task costs $0.87 on Claude and $0.26 on GPT, and the weekly capacity of each plan becomes a count of tasks:

“Fee per task” is the weekly share of the subscription (price ÷ 4.33) divided by tasks per week. Every row is two orders of magnitude below the API price of the same task, so every plan beats paying per token for anyone who uses it seriously.

Does ChatGPT Pro 20x really give 20x the weekly Codex usage?

OpenAI has not said so in writing, and it is the one number that could flip this study. Codex keeps its lead at $200 as long as the real weekly multiplier is above about 7x.

The Codex Pro rows rest on one assumption: that the Pro 20x weekly window really holds twenty Plus weeks. OpenAI publishes 20x for the 5-hour window. Anthropic's experience, where 20x per session became about 2x per week, is a warning that the two need not match. So instead of assuming, we can ask how large the real weekly multiplier m must be for Codex to win.

Codex tasks/week = m × $67 ÷ ($0.87 ÷ r), where r is how many times more Claude spends per task.

If OpenAI's weekly allowance scales even a third as far as its label, ChatGPT Pro 20x delivers more finished work than Claude Max 20x. If it scales the way Anthropic's did, at 2x, Claude wins comfortably. You can check this in your first week on Pro. Watch the weekly gauge that Codex already shows you: if a full Plus-sized day of work moves it by about 1 point, you are getting the full 20x.

Is a $200 AI coding subscription cheaper than the API?

Every one of these plans beats paying per token, and by a wide margin. A $200 subscription pays for itself at $46 of API-equivalent usage a week, and a $100 plan at $23. That is roughly two hours of steady Opus agent work, or one heavy afternoon. Our user went past that line on the first Monday of every week.

Claude Max or ChatGPT Pro: which should you buy at $100 or $200?

One availability note: buildthisnow reports that new sign-ups to ChatGPT's $200 Pro tier have been paused since 10 September 2026. Check before planning a rollout around it.

Is it better to split $200 across Claude Max 5x and ChatGPT Pro 5x?

It beats Claude Max 20x, but not ChatGPT Pro 20x. Several guides now recommend buying the two $100 plans instead of one $200 plan. On our numbers that split holds at least 2,687 tasks a week, against 2,138 on Claude Max 20x and 5,154 on ChatGPT Pro 20x.

The split wins on things the table does not show. You get both model families, so the hard refactor goes to Opus and the ticket queue goes to Sol. When one vendor’s limit runs out mid-afternoon, the other still has its own. And if Pro 20x’s weekly window turns out to scale far less than 20x, the split is the safer bet, since it doesn’t depend on that multiplier nearly as much.

For FinOps teams

• Book seats as tooling, not COGS. A flat seat is a fixed cost with an enormous implied discount. Once seat usage leaks into anything customer-facing, you lose unit economics, because none of it can be tagged to a request, a tenant or a feature.
• Measure the ratio, not the label. Both logs sit on the developer's laptop. A weekly script that turns them into API-equivalent dollars per seat is the cheapest utilisation report you will ever build, and it shows who should move down a tier.
• Right-size downwards first. On Claude, the gap between 5x and 20x per week is 1.5–2x for 2x the price. Move a seat up only when it logs repeated weekly-limit hits.
• Keep two vendors. Limits change without notice, as the end of the 150% promo showed. A $20 second seat is the cheapest insurance against a week of blocked developers.

Limitations

• One subject. The logs come from one heavy user whose work is mostly agentic coding. Your cache share, model mix and session length will differ, and they are what set your ratio.
• Codex Pro is projected. We measured Plus windows directly. Pro 5x and Pro 20x are multiples of those windows, not observations.
• The Claude cap moves with model mix. The same metered week reads as more dollars on Opus 5 than on Opus 5.5, which has a cheaper price list. We priced each turn at the model that actually ran it.
• “Tasks” is one yardstick. A Tetris clone is a reasonable proxy for a mid-sized feature. It is not your codebase.
• API-equivalent is not cash. It is what the same tokens would have cost on the metered API. Nobody would have run this much traffic at list price, so it overstates the real saving.

Sources

• Anthropic: What is the Max plan?
• Anthropic API pricing
• OpenAI API pricing
• OpenAI: Codex plans and usage limits
• codeongrass: Claude Code vs Codex for heavy users
• buildthisnow: Codex vs Claude Code usage limits
• DataCamp: GPT-6 Sol vs Claude Opus 5.5
• Finout: Claude Code pricing 2026
• OpenRouter model price list

Related

• Claude Max 20x vs 5x: the weekly limit, measured
• How much Codex ChatGPT Pro includes
• How many tokens Claude Pro and Max give you
• SaaS AI credits and metering
• Committed spend discounts
• Proving AI ROI

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research