Industry analysis · July 24, 2026

Codex Pro usage cut: the July report and what changed since

A user's measurements suggest the $100 Codex plan fell from $675 to $600 in weekly API-equivalent usage. The evidence is credible, but not yet universal.

Reading time
10 min
Checked
Nov 19, 2026
Editorial abstraction of a shrinking measured allowance inside a fixed calibration frame
A fixed subscription price needs a consistent measuring stick.
Bottom line

The July 11% decline was one account’s estimate, not a published quota. OpenAI now reports a higher Pro 100 allowance. Compare completed work under consistent settings before deciding whether your plan has become worse value.

A usage bar can tell you how much Codex you have left without telling you how much work that buys. That was the useful discovery behind July’s reported 11% allowance cut, and it still matters after OpenAI’s latest increase.

Updated November 19, 2026. OpenAI now says Pro 100 has twice its previous Work and Codex allowance. The July figures below remain a historical, single-account measurement; they are not today’s quota. We rechecked the linked sources for this update, but did not reproduce the user’s experiment or measure a paid account.

In the original July 23 report, one user’s estimated weekly allowance fell from roughly $675 to $600 in API-equivalent usage. That is an 11.1% decline. The interesting question is what changed behind the percentage meter, not whether those dollar figures were an official entitlement.

My advice: keep the historical report as a reason to measure your own workflow. Don’t use it as a buying guide for the current plan.

How the user measured the change

The original Reddit post is more rigorous than the usual “my limit feels worse” complaint.

The author, RealSuperdau, tracked the $100 Pro tier they called Pro Lite. On both July 13 and July 23, they ran heavy Codex tasks and repeatedly recorded two values:

  1. the percentage of the weekly allowance still showing in Codex;
  2. the API list-price equivalent of the input, cached input, and output tokens reported by ccusage.

They then fitted the observations to estimate the cost-equivalent value at which the weekly meter would reach zero.

Test dateMeasurementsObserved sampleEstimated full weekly allowance
July 1320+About 28% of the meter and $193 API-equivalentAbout $675
July 2320+About 16% of the meter and $96 API-equivalentAbout $600

The author says alternative regression methods designed to account for Codex rounding its meter to whole percentages produced estimates within about $5 of each other.

They also tried to keep the workload comparable, using GPT-5.6 at high through ultra reasoning in non-fast mode, with occasional automatic fallback to GPT-5.5. That does not eliminate every variable, but it makes the result more useful than a comparison based on time spent in the app or number of prompts.

Why “API-equivalent” is not OpenAI’s actual cost

The $675 and $600 figures do not mean OpenAI paid those amounts to serve the account.

They represent what the recorded token mix would have cost a customer at published API rates. Retail API prices do not reveal OpenAI’s underlying compute cost. Subscription metering and API billing also need not use the same weights.

The comparison is useful only to the extent that its measuring stick stays constant. A changed model mix, rate table or relationship between token use and the meter could change the estimate. That is why comparable settings matter as much as collecting more samples.

One interpretation is a smaller discount relative to API list prices, rather than evidence of OpenAI’s actual subsidy. At the earlier estimate, a power user could consume about 6.75 times the monthly subscription price in API-equivalent work every week. At the later estimate, that multiple fell to about 6 times.

Under the author’s assumptions, the subscription looked heavily discounted relative to buying those measured tokens through the API at the time. The complaint is not that $600 of weekly API-equivalent work is poor value for $100 a month. It is that the included value may be shrinking without a disclosed change to the product.

What OpenAI officially promises now

The Pro tiers help page, checked November 19, lists Pro 100 at $100/month. It says Pro 100 and Pro 500 received double their previous Work and Codex allowance automatically. For Pro 200, the lower allowance rose to match the higher one; the previously announced reduction was canceled. These are vendor statements, not our account measurements.

That supersedes this article’s earlier reliance on a 5x-Plus promise. It does not establish a new API-equivalent dollar quota, and doubling July’s $600 estimate would mix two different baselines.

Speed settings also matter. The same help page says Fast consumes included usage at 2.5 times Standard’s rate. For supported models, Ultrafast consumes included usage at 8 times Standard, while purchased credits use a different multiplier. A larger allowance does not guarantee proportionally more finished tasks.

The Codex pricing page still says local messages and cloud chats share an allowance and that weekly limits can apply. Its model-specific ranges are estimates, not a fixed number of successful jobs.

The practical improvement is that OpenAI now describes allowance changes and speed multipliers explicitly. The remaining gap is a stable, exportable baseline that lets a subscriber compare like-for-like work over time.

Why one account is not a global audit

One user’s regression is evidence, not a global audit.

The author explicitly acknowledges that limits might differ between accounts or be part of an A/B test. The test windows also used different portions of the weekly meter, and Codex can change the model used during a task. An API-equivalent calculation may not perfectly match the internal weights OpenAI applies to different models, reasoning levels, speed settings, cached tokens, or tool use.

There is also a June operational precedent that should not be confused with a deliberate plan cut. In late June, OpenAI investigated reports of Codex allowances depleting too quickly and said some accounts had been incorrectly rate-limited by abuse and fraud systems. OpenAI marked that incident resolved on June 29 and said it had not observed broader degradation.

That incident does not explain a July 13-to-July 23 decline. It does show that a meter draining faster can result from a bug or account-specific enforcement rather than an intentional commercial decision.

The defensible historical conclusion is narrow: the author’s estimate fell by about 11%. Neither that experiment nor today’s documentation establishes the cause of the July change.

The wider subscription problem

Software subscriptions used to be easy to understand. You paid for a feature set, storage capacity, number of seats, or a clearly stated transaction allowance.

Agent subscriptions sell something less stable: a claim on future inference. The cost of fulfilling that claim changes with the model, reasoning depth, context length, tool calls, caching, and the user’s ability to run multiple long jobs. Providers want the simplicity of a flat monthly price without committing to a fixed amount of the expensive resource underneath it.

That creates a black-box contract:

  • the price is exact;
  • the advertised multiplier is relative;
  • the real allowance is hidden;
  • the meter is rounded;
  • and the vendor can change models or internal weights.

A reduction does not need to appear in the plan name or monthly bill. It can arrive as a task that used to consume 3% now consuming 4%, a weekly reset that lasts five days instead of six, or a model that is quietly weighted more heavily against the same percentage bar.

For individual users, that makes comparison shopping unusually difficult. For businesses building workflows around subscriptions, it makes capacity planning fragile. A workflow that fits comfortably within a plan this month may require top-up credits next month even when its prompts and output remain unchanged.

What OpenAI should publish

OpenAI does not need to guarantee unlimited use. It does need to make paid allowances legible.

A useful disclosure would include:

  • the fixed weekly credit or compute allowance for each plan;
  • the rate at which each model, reasoning level, and speed setting consumes it;
  • a dated history of allowance or weighting changes;
  • account-specific restrictions clearly separated from the standard plan limit;
  • unrounded usage data that customers can export;
  • advance notice when the effective included allowance is reduced.

OpenAI already has a credit rate card for usage beyond included limits. Extending that clarity to the subscription allowance would let users distinguish a policy change from a bug, a more expensive model, or an unusually heavy task.

Check your own allowance without wasting it

Use work you already need to do. Burning through a subscription to prove that it is running out defeats the point. This is a proposed tracking exercise, not a benchmark we ran.

  1. Record the starting conditions. Save the date and timezone, plan name, client version, selected model, reasoning level and speed mode. Note whether you are using included allowance, purchased credits or API billing.
  2. Read the right meter. Open your usage dashboard, or enter /status in an active Codex CLI session. Record each displayed allowance and its reset time separately. A context-window percentage is not a subscription balance. OpenAI documents these usage checks in Using Codex with your ChatGPT plan.
  3. Choose a bounded job. For example, ask for a fix to one failing test on a recorded commit. Decide in advance what counts as accepted: the test passes, relevant regressions pass and you approve the diff. Keep the prompt and starting revision with the result.
  4. Record the finish. Save the meter again, retries, review minutes and whether the work was accepted. Note other account activity during the interval. Do not attribute a shared meter’s entire movement to one task while other sessions are running.
  5. Compare several comparable jobs. Keep settings consistent and separate observations that cross a reset, tier change or announced allowance change. Treat model fallback as a changed condition. If the display is rounded, a one-point movement is too coarse to support a precise claim about capacity.

A small private log is enough:

date_timezone,plan,client_version,model,reasoning,speed,billing_mode,
starting_commit,task,window,reset_time,remaining_before,remaining_after,
other_activity,retries,review_minutes,accepted,notes

Record percentage changes in percentage points. As an invented example, a fall from 80% remaining to 76% is four points of the displayed allowance. It is not evidence that the task cost four dollars. Keep API-equivalent estimates in a separate column with their rate date and model mix.

If access stays blocked after the displayed reset, retain the timestamps and error text for Support. OpenAI says Support can investigate incorrect metering or access that fails to return, but cannot reset limits. Avoid including private repository contents in a support report.

What subscribers can conclude

July’s experiment is a useful example of measuring an opaque product. November’s announced increase is a reason to start a fresh comparison. Neither tells you how many reviewed fixes your own subscription will deliver.

For a renewal decision, I would track cost per accepted deliverable over one consistent period: the subscription cost allocated to that period, plus paid extras and review time valued at your own hourly rate, divided by accepted results. If nothing is accepted, report that directly rather than dividing by zero. Keep job sizes comparable; one typo fix and a feature migration are not interchangeable units.

A cheaper model that needs three repair attempts may be the expensive option. A faster mode may be worth its heavier meter use when it saves meaningful waiting time. The model-routing lesson develops that decision: spend on the work that needs it, and judge the result after review.

Sources

Checked November 19, 2026. The Reddit post is the original author’s account of the July experiment; OpenAI links describe vendor policy or incident history.

Put this to work

Track AI coding spend as cost per accepted deliverable rather than tokens, messages, or advertised multipliers.

Try

For one week, log the task, plan used, retries, review minutes, and whether the result shipped.

Prove it worked

Divide the total plan and review cost by accepted deliverables, then compare that number with a second model or workflow.

Where it can pay

Freelancers and small teams can protect project margin by pricing from measured delivery cost and keeping a failure allowance.

Keep in view

  • A subscription allowance is useful only when you connect it to completed, accepted work.
  • Retries, failed runs, review time, and model switching can cost more than the headline plan price.
  • One user's measurements can reveal a risk without proving a universal policy change.
Learn the workflow: routing models by cost and reliability