Were Codex Limits Reduced? What the Evidence Shows

OpenAI has acknowledged that GPT-5.6 Sol can consume limits faster than expected, while user reports, client bugs and tracker estimates show conflicting results. A secret systemwide reduction remains unproven.

Verdict: SUPPORTED BUT UNPROVEN. Reports that some Codex users are exhausting their allowance unexpectedly fast are credible and, in several cases, technically detailed. The stronger allegation—that OpenAI quietly cut the underlying weekly allowance for all or most subscribers, sometimes by 50% or more—is not established by the available evidence.

That distinction matters. A smaller allowance and faster consumption under an unchanged allowance look identical in a percentage meter. So can a delayed meter update, a plan downgrade, a reset that moves the weekly window, subagents running additional turns, or a client that repeatedly resends a large context while waiting for a command. OpenAI does not expose enough historical, per-turn account data to separate those causes cleanly for every report.

The complaint wave itself is real. At the reporting cutoff, an August 20 X post by Alex Getman alleging that a Plus allowance had fallen from roughly $160 to $80 in API-price-equivalent value had 193,833 views, 1,858 likes, 210 reposts and 172 replies. A second widely shared post, which combined the Plus estimate with a claimed 77% Pro 5× decline, had 90,946 views and 1,049 likes.

Those figures demonstrate reach, not truth. Both posts ultimately depend on a handful of account-level estimates rather than controlled fleet telemetry. In fact, Getman’s own follow-up presented contrary data: another tracker estimated a Pro 20× allowance at $2,043 per seven days and reported no significant recent decline.

What is confirmed

Claim	Finding	Evidence
Codex is included with ChatGPT plans	Verified	OpenAI pricing lists Codex and ChatGPT Work as included in Free, Go, Plus, Pro, Business, Edu and Enterprise. It does not list a separate “Codex subscription.”
Subscription and API billing are separate	Verified	OpenAI authentication documentation distinguishes ChatGPT sign-in for subscription access from API-key sign-in for usage-based access.
Codex and ChatGPT Work share usage	Verified	OpenAI says they use the same pricing, credits and usage limits. Current pricing also identifies ChatGPT for Excel as sharing limits on Plus and Pro.
Published usage is workload-dependent	Verified	OpenAI says model, context, reasoning, tools, retrieval and caching all affect consumption; additional weekly limits may apply.
Sol produced unexpectedly high burn for some users	Verified	On July 28, OpenAI’s Tibo said Sol was using limits faster than expected for some power users and described efficiency changes intended to make typical use last about 18% longer.
A current blanket quota cut has been announced	Not supported	No such change appears in the current pricing page, changelog or status history reviewed for August 1–20.
A current systemwide quota incident is acknowledged	Not supported	The OpenAI status page reported full operation at the cutoff. Status is aggregate and does not exclude individual-account problems.

What subscription is the Codex app actually using?

There is no separately documented “Codex subscription.” The consumer product is access to Codex included with a ChatGPT plan. Pro presently offers 5× or 20× the Plus usage level, starting at $100 per month. Users can then purchase ChatGPT credits after exhausting included usage.

API-key use is a separate route. OpenAI’s documentation says a user signs in with ChatGPT for subscription access or with an API key for usage-based access. In the desktop app, API authentication requires choosing another sign-in method and entering a key. In the CLI, codex login --with-api-key is an explicit operation. OpenAI documents no silent fallback from a ChatGPT subscription session to API billing.

On the locally examined account, codex login status returned Logged in using ChatGPT. Numeric rate-limit metadata identified a Pro plan and the Codex weekly window, with no purchased credits. That session was therefore drawing from the Codex/shared agentic allowance included in ChatGPT Pro—not from separately metered Platform API usage.

Users can confirm their own route in the desktop profile menu or with codex login status. The account’s usage dashboard and the CLI /status command show the remaining subscription allowance. API-key users should instead verify charges in the OpenAI Platform usage dashboard.

Why similar-looking tasks can cost very different amounts

OpenAI publishes estimated message ranges rather than a fixed number of tokens per subscription. For Plus, the current examples range from 10–100 local Sol messages per rolling five hours, 25–200 for Terra and 250–2,000 for Luna. Pro 5× and 20× multiply those ranges. Local messages and cloud work share the five-hour window, and OpenAI warns that additional weekly limits may apply.

The underlying accounting is now visible as model-specific credits per million tokens. Sol costs 125 credits per million uncached input tokens, 12.5 for cached input and 750 for output. Terra’s corresponding rates are 50, 5 and 300; Luna’s are 5, 0.5 and 30. OpenAI says a GPT-5.6 message averages 5–40 credits, an exceptionally wide range.

Several settings increase consumption further. Fast mode makes GPT-5.6 and GPT-5.5 run approximately 1.5× faster while consuming credits at 2.5× the standard rate. Image generation uses included limits about three to five times faster on average. Higher reasoning uses more tokens. Ultra mode launches subagents, which can turn one visible request into several model runs. Every enabled MCP server adds tool descriptions and context. Prompts, attached files, chat history, tool results, instructions and outputs all count.

OpenAI effectively acknowledged this problem on July 28. In a 3.1-million-view post, Tibo said the company had not reduced subscription usage, but found that Sol worked longer, made more tool calls and coordinated more complex workflows than earlier models. High reasoning on Sol could use more than High on GPT-5.5. Programmatic tool calling generated more responses and cached input, while waiting for tools and running many web searches proved particularly expensive. OpenAI said median usage looked efficient, but it had missed a long tail of power users.

That official explanation predates the August 20 wave and cannot rule out a later change. It does establish, however, that unexpectedly fast effective burn can occur without a smaller plan allowance.

The trackers disagree

The viral “halved” figure comes from NerfTrack, a new, local-only open-source estimator. It prices local token events at API rates, pairs the token-cost increase with a positive change in the rounded weekly-used percentage, and extrapolates the value of 100% of the window. Its headline is the median of the latest seven valid cumulative estimates.

That is a serious attempt at measurement, but it is not a direct reading of OpenAI’s internal allowance. Shared activity from another Codex surface can move the server percentage without appearing in the same local token log. Meter updates can arrive late. The displayed percentage is rounded. Reset timestamps can change. Tool-unit and non-text charges may be missing. Repricing historical tokens with a current rate card can also make old and new periods look more comparable than they were. The methodology explicitly says its dollar figure is an API equivalent, not a ChatGPT bill.

TiboTattle uses another fitting method and currently tells the opposite story for Pro 20×. Its August 20 public estimate was $2,043 per seven days, with a plausible middle-80% range of $1,453–$2,394. The central series sat around $1,971–$1,983 from August 8 through August 18 before rising to $2,043 on August 19. That is inconsistent with a new 50% systemwide Pro 20× cut.

But TiboTattle’s apparent stability is not definitive either. The current public series came from one contributing account and 17 qualifying reset fits in the trailing 30 days. The strongest conclusion from the two trackers is therefore not “cut” or “no cut.” It is that current public samples are too small and method-dependent to establish a fleet-wide answer.

Bug reports reveal multiple mechanisms

A GitHub API search found 29 issues created from August 17 through August 20 carrying the rate-limits label. That number is a measure of reporting activity, not 29 demonstrations of a smaller quota: the label covers limits, entitlements, reset behavior, capacity and token reporting.

Several reports are nevertheless concerning. Issue #39699 describes more than twice the apparent burn for recent Sol sessions even though 96%–98% of input was cached. Issue #39763 reports 6% of a Pro 20× weekly allowance in 90 minutes, while acknowledging that the user’s tiny visible prompt arrived with system instructions, workspace rules, environment metadata and tool catalogs. Issue #39796 reports a Pro allowance falling from roughly full to 5% in four hours alongside 36 GB of memory use. These issues include feedback identifiers or version data but no public server-side accounting response.

The most useful evidence is forensic rather than rhetorical. A comment on issue #38335 traced one session to 90 polling turns while a long command ran. Each poll resent roughly 240,000 input tokens; the loop accumulated about 21.6 million input tokens while producing almost no output. That is a concrete client-workflow mechanism capable of burning substantial allowance without any change to the underlying cap.

Another dramatic case turned out not to be a usage bug. Issue #39167 initially documented a jump from 88% used to 100% during a nine-hour period with no client requests. The author later found that the subscription had renewed from Pro 20× to Pro 5× in the middle of the weekly window. The numerator did not change; the denominator became roughly four times smaller. The author closed the report, while retaining two valid product criticisms: the meter lagged, and the dashboard classified subagent traffic as CLI activity.

There is also real precedent for account-specific failures. From June 26–29, OpenAI investigated unexpectedly fast Codex depletion and concluded that abuse and fraud systems had incorrectly rate-limited certain accounts. OpenAI said the impact was limited and that it had not observed broader degradation.

The local metadata audit

One ChatGPT-authenticated Pro account was examined retrospectively using only numeric token and rate-limit metadata; no prompt or response text was read or emitted. Token deltas were priced with OpenAI’s current credit rate card and compared with the server-reported weekly-used percentage. The resulting dollar figures are normalization estimates, not charges or official allowance values.

The two latest reset identities spanning 99 percentage points produced rough API-equivalent estimates of $2,434 and $2,865 per week. They overlap and therefore are not independent weekly observations. The current window, measured from 0% to 28% used, extrapolated to approximately $2,519. Earlier August windows ranged from roughly $1,220 to $3,420.

Those results are too inconsistent to support an exact allowance claim. They are affected by overlapping reset identifiers, promotional or manual resets, concurrent work, rounded percentages, missing tool charges, unpriced events and the omission of the first cumulative sample in each local file. The audit’s narrow finding is still useful: this account does not show a current collapse to half of its two closest full-window estimates.

A prospective fixed-task benchmark was not run because no consumption cap was authorized. Even if run, a present-day experiment could measure current variance but could not prove a historical reduction without equivalent older runs.

A timeline that can make today feel smaller

February 2: OpenAI launched the Codex app and said paid plans would receive double rate limits “for a limited time.” The current pricing page no longer describes paid limits as doubled. The changelog does not give a precise expiration date.
June 11: OpenAI introduced banked rate-limit resets and referral rewards, temporarily increasing some users’ effective capacity.
June 26–29: OpenAI confirmed incorrect rate-limiting for a limited set of accounts and resolved the incident.
Mid-July: OpenAI temporarily removed the five-hour limit for Plus and Pro while soliciting feedback.
July 28: OpenAI denied reducing subscription-plan usage, reset users’ limits, announced Sol-efficiency improvements and said the five-hour limit would return.
August 7: Codex CLI 0.147.0 shipped multiple workflow and plugin changes. Its release notes did not announce a quota reduction.
August 18: CLI 0.148.0 added estimated per-thread credits or cost to eligible /status surfaces. Again, no general limit cut was announced.
August 20: The current complaint wave accelerated across X, Reddit and GitHub. OpenAI’s status page listed no quota-depletion incident at the reporting cutoff.

Users comparing today against a double-limit promotion, repeated free resets or a temporarily absent five-hour cap are comparing against an unusually generous high-water mark. That can produce a genuine feeling of reduction even if the current base allowance has not just changed. It does not explain every account-specific report.

Best-supported explanations, ranked

Model and workflow consumption increased. This has the strongest direct support: OpenAI acknowledged Sol’s long-tail inefficiency, while technical traces show subagents, large context and polling loops can multiply hidden work.
Shared usage and expensive settings obscure attribution. ChatGPT Work, local and cloud Codex, and specified agentic features can share limits. Fast mode, images, Ultra and high reasoning materially raise consumption.
Promotional baselines ended or changed. Double limits, banked resets and the temporary removal of the five-hour cap made earlier periods more generous. The exact double-limit expiry remains undocumented.
Meter, reset, entitlement or account-specific bugs exist. The June incident proves precedent, and current GitHub reports contain plausible unresolved anomalies. No current broad incident has been confirmed.
A new systemwide secret cut. Possible in principle because OpenAI does not publish the absolute weekly budget, but not demonstrated by current public evidence. The available tracker and local-account results conflict with it.

What users can check now

Run codex login status to distinguish subscription access from API-key billing. Record the usage dashboard and /status before and after a task. Note the model, reasoning effort, Fast setting, app/CLI version, subagent count and enabled MCP servers. Check for cloud work, scheduled tasks, ChatGPT Work or Excel activity, voice-started tasks, images and plan renewal changes. Treat the percentage as a lagging, rounded account-level meter rather than a per-thread invoice.

For routine work, Terra or Luna can extend usage substantially. Disable unneeded MCP servers, reduce oversized AGENTS.md instructions, start a fresh thread when context has become enormous, and avoid repeated polling around long silent commands. Those steps will not repair a backend entitlement bug, but they remove major documented sources of unexplained burn.

OpenAI response and what remains unknown

No fresh outreach was sent because external contact was not authorized. OpenAI’s latest directly relevant public response located for this investigation is the July 28 statement denying plan reductions while acknowledging unexpectedly high Sol consumption. OpenAI has not publicly answered whether any unpublished weekly budget, staged experiment, enforcement weight or cache-accounting rule changed during the August 17–20 period.

OpenAI also does not publish the absolute weekly token or credit budget for each individual plan. Users cannot retrieve a complete historical ledger allocating every percentage point among threads, subagents, retries, tool calls and shared products. The precise end date of the February double-limit promotion is unclear. The affected population, if any, cannot be estimated from algorithmically amplified social posts or voluntary tracker samples.

The responsible conclusion is therefore narrower than either side of the online argument: some Codex users are encountering real, sometimes severe effective depletion, but the evidence does not presently establish that OpenAI quietly halved Codex limits systemwide. A definitive claim of a new cap reduction requires controlled multi-account before-and-after data or an on-record OpenAI confirmation.

Methodology

Four original X report/counter-report posts, four Reddit threads and six GitHub issues were reviewed, for 14 public-report items, alongside official documentation, status history and two independent tracker methodologies. Reposts and quoted repetitions were not counted as new evidence. Thirteen of the 14 items named a plan and/or model and supplied a numeric observation somewhere in the post or thread; only two public GitHub cases contained enough detail to diagnose a specific mechanism. None provided a replicated, cross-account, fixed-task before-and-after benchmark.

X engagement was recorded at approximately 2:20–2:30 p.m. PDT on August 20. X’s recommendation algorithm, follower graph and viral reposting make this sample unsuitable for estimating prevalence. Reddit timestamps and rankings are dynamic. GitHub issue labels reflect reporter and triage classification rather than confirmed defects.

The local audit read only numeric token_count and rate-limit metadata. It applied the current August 20 credit card retrospectively, excluded the first cumulative token sample in each file, and did not price unknown/Spark events or non-text tool units. Its estimates should not be used as invoices or contractual allowance values.

Compact claim ledger

Material claim	Classification	Evidence	Residual uncertainty
A substantial complaint wave exists	Verified	X engagement snapshots, Reddit threads, 29 newly created rate-limit-labelled GitHub issues	No population denominator
Some users see severe effective depletion	Corroborated	Independent X, Reddit and GitHub reports across Plus and Pro	Workloads and account states differ
OpenAI reduced every subscription allowance	Unknown / not supported	No official announcement; conflicting tracker and local results	Absolute weekly budgets are unpublished
Sol can consume limits faster than expected	Verified	July 28 OpenAI statement	Does not quantify every workload or later period
Fast mode and images increase consumption	Verified	Current OpenAI pricing and speed documentation	Actual task totals remain workload-dependent
ChatGPT Work and Codex share usage	Verified	Current OpenAI pricing	Ordinary non-Work ChatGPT conversations are not thereby proven to share Codex usage
API-key use is separate from subscription use	Verified	OpenAI authentication documentation; local login status	Account-specific route must be checked
A client polling loop can cause extreme burn	Reproduced by reporter	GitHub #38335 trace	One documented session, not fleet prevalence
“Idle drain” always indicates hidden billing	Contradicted	GitHub #39167 was caused by a plan downgrade	Other idle/reset anomalies remain unresolved
The examined local Pro account shows a 50% current cut	Not reproduced	Current-window estimate near two recent full-span identities	One account; estimator has substantial uncertainty

Publication gate

Gate	Result	Note
Billing and authentication taxonomy verified	PASS	Current OpenAI documentation and local login status agree.
Central conclusion supported by two independent streams	PASS	Official documentation/statements plus public trackers, GitHub traces and local metadata.
Historical change checked against dated records	PARTIAL	Promotion and reset history are documented; exact double-limit expiry is not.
Social sampling and limitations disclosed	PASS	Original-item count, engagement cutoff and selection bias are stated.
Controlled prospective benchmark completed	FAIL	No consumption cap was authorized; current testing cannot recreate historical conditions.
Fresh OpenAI response obtained	FAIL	Outreach was not authorized.
Material contradiction preserved	PASS	Conflicting trackers and counterexamples are central to the article.
Headline no stronger than evidence	PASS	The chosen headline rejects a proven blanket-cut claim while recognizing real problems.

Publication decision: Publishable as a sourced “what the evidence shows” investigation. Not publishable as confirmation that OpenAI secretly halved Codex limits.

Direct sources

All sources retrieved or rechecked August 20, 2026.

OpenAI and official project sources

Trackers and community reports

Corrections note

This article uses live documentation and engagement snapshots that can change. Any subsequent OpenAI statement, tracker sample expansion, corrected issue diagnosis or controlled benchmark that materially changes the conclusion should be added with a dated correction. API-equivalent dollar estimates must never be relabelled as subscription charges or guaranteed allowance values.