Claude Max 20x isn’t 20x: 60 days of logs vs ChatGPT Pro

Updated: September 27, 2026

By the LLM CFO team

At $200 a month, ChatGPT Pro 20x (Codex) buys about 2.4 times more finished coding work than Claude Max 20x. At $100, Claude Max 5x is the better seat. Every one of these plans costs less than paying for the same use through the metered API. We priced 60 days of real usage logs to show why, and what to check before signing off.

The request usually arrives in a single Slack message: “Can I expense a $200 AI seat?” Sometimes the seat is Claude Max, sometimes ChatGPT Pro for Codex, and more and more often it is both. The numbers on the two pricing pages look like they can be compared directly. They can't. This is the analysis a budget owner needs before approving either one, or before deciding which one becomes the team standard.

The decision in one table

“API value” is what the same tokens would have cost on each vendor's pay-per-use API at list price. A “task” is a mid-sized coding job; the unit comes from an independent benchmark described below. “Cost per task” is the weekly share of the fee divided by tasks per week.

Why the two “20x” labels are not comparable

Both vendors print “20x” on the $200 tier, but they are multiplying different things.

• Anthropic's 20x is per 5-hour session. A separate weekly cap sits on top of it, and Anthropic does not publish its size. For a heavy user the weekly cap is the constraint that binds. Independent testing and our own logs both put the Max 20x weekly allowance at only 1.5 to 2 times the Max 5x one, for twice the price.
• OpenAI's 20x is published as usage ranges that scale by exactly 5x and 20x from Plus, and extra Codex credits are billed at API parity. OpenAI does not say in writing whether the weekly window scales by the same factor. That gap is the one open question in this analysis.

For a budget owner, the conclusion is simple. Don't pay for a multiplier. Pay for weekly capacity, and measure it.

What one heavy user actually consumed

Both coding agents keep complete local logs. We re-priced every logged request at the API price in force on the day it ran. Claude Code also records the moment a limit is hit, which lets us put a dollar value on the cap itself.

Our user hit the Max 20x weekly cap twice in August, after $2,085 and $2,379 of usage. In a later week they reached $2,636 without a hit. So the cap is not a fixed dollar figure: Anthropic meters in its own units, and pricier models use them up faster. All of these weeks fell inside a promotion that raised limits to 150%. Scaled to today's 125%, the Max 20x weekly cap is worth about $1.74k to $2.2k of API usage, with a central estimate of $1.86k. On Max 5x, the weekly cap never bound. What stopped our user was the 5-hour session limit, which cut in after a median of $37 of work.

On Codex, a full weekly window on the $20 Plus plan has been worth a median $67 since August. It was $115 before that. The quota did not shrink: the default model moved to GPT-5.6 and GPT-6 Sol, which cost less than half as much per token, so the same compute is now worth fewer dollars at list price.

For context, Finout estimates that heavy Claude Code users generate $600 to $1,500 of API-equivalent usage a month. Our subject is well above that, so read these numbers as the ceiling of what a seat can deliver, not the average.

The line item nobody forecasts

In August, output tokens accounted for $1,038 of our user's $6,986 Opus usage. The rest was context the agent re-read and cached on every step: $3,510 of cache reads and $2,437 of cache writes. Cache reads made up 96% of the tokens.

This matters most when a team moves off seats and onto the API. A budget sized from “how much code does it write” will come in about seven times too low. Any chargeback model for coding agents needs cache reads and cache writes as their own lines.

From dollars to finished work

The dollar value of a plan measures what the vendor gives up, not what your team gets. The two model families spend very differently to reach the same result.

Claude still leads on quality benchmarks. On Terminal-Bench 4.0, Opus 5.5 scores 66.4% and GPT-6 Sol 60.5%. That lead is worth paying for on hard, ambiguous engineering. It does not justify a threefold cost gap on routine work. Converting each plan's weekly value into tasks gives the ranking below.

The one assumption to verify

The ChatGPT Pro figures are projections: measured Plus weeks multiplied by OpenAI's published 5x and 20x. If Pro 20x's weekly window turns out to scale less than its 5-hour window, as Anthropic's did, the lead shrinks. The chart shows how far it can fall before Claude wins.

Codex stays ahead at $200 as long as the real weekly multiplier is above about 7x, or 6.6 to 8.4x depending on the benchmark. That is roughly a third of the label. Check it in the first week of a pilot: Codex shows a live weekly gauge, and a full Plus-sized day of work should move it by about one percentage point on Pro 20x.

Break-even: when does a seat pay for itself?

Any engineer who uses a coding agent daily clears these thresholds before Monday lunch. The question is never whether a seat is cheaper than the API; for internal use, it always is. The real question is which seat, and whether the $200 tier is justified over the $100 one.

A budgeting policy you can adopt as written

1. Default: Claude Max 5x at $100 per engineer who uses agents daily. It gave the most headroom per dollar in our data.
2. Upgrade on evidence, not request. Move someone to a $200 tier only after two consecutive weeks of logged limit hits. Both agents record these locally, so the evidence costs nothing to collect.
3. High-volume ticket work: ChatGPT Pro 20x at $200, the lowest cost per finished task, pending the weekly-multiplier check above. Note that buildthisnow reports new $200 Pro sign-ups paused since 10 September 2026.
4. Resilience: a $20 ChatGPT Plus seat next to a Claude seat. Limits change without notice, and a second vendor with its own quota costs less than one blocked afternoon.
5. Accounting: book seats as fixed tooling expense. Anything customer-facing goes on the metered API, where every request can be attributed to a product, a tenant and a margin line.

Caveats

• This analysis rests on one heavy user's agentic coding logs. Your team's cache share, model mix and session habits will shift the numbers.
• ChatGPT Pro figures are projections from measured Plus windows, not observations.
• API-equivalent value is not cash saved. It is what the same traffic would have cost on the metered API.
• A benchmark task is a proxy. Your codebase is not a Tetris clone.

The full methodology, the pricing formula and the week-by-week log tables are in the FinOps edition of this study on finopsllm.com.

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

Related

• I used Claude Max 20x heavily for a week
• Claude Max 20x vs Qwen 3.8 token plan
• LLM cost visibility for CFOs
• Cost per profitable outcome