The Reason People Switched to Codex Comes Back Tomorrow

OpenAI is restoring the rolling five-hour usage limit on ChatGPT Work and Codex for Plus subscribers, announced by engineering lead Thibault Sottiaux on August 25, 2026 and taking effect the next day. The cap was removed on July 12, leaving roughly six weeks where only a weekly allowance applied. Pro at $100 and $200 stays exempt for the coming months. The two limits operate independently, so hitting either one pauses your usage until that specific limit resets. His stated reasons are smoothing compute load plus stopping casual users burning a week of quota in one session. Anthropic’s Sonnet 5 introductory pricing ends September 1, six days later. Best for understanding what your subscription actually gives you this week. Not ideal for anyone hoping this is the last change.

On July 12, OpenAI removed the five-hour usage cap for Plus subscribers on ChatGPT Work and Codex. Only the weekly allowance remained.

Some developers moved to Codex specifically because of that. No five-hour window meant a long session could run without hitting a wall mid-task, which is the single most irritating thing about agent subscriptions.

Tomorrow the cap returns.

Six days after that, on September 1, Anthropic’s Sonnet 5 introductory pricing ends and the rate moves from $2 and $10 per million tokens to $3 and $15.

So both options change inside a week, in the same direction.

What Was Announced

Verified across eight sources including the original announcement.

The important mechanical detail is that the two limits work independently. Reaching either the five-hour window or the weekly allowance pauses included usage until that particular limit resets, as Studio Global’s writeup sets out. Your weekly quota having plenty left does not help if the five-hour window is spent.

Sottiaux gave two reasons. The cap smooths load across OpenAI’s compute, which lets the plan stay generous on weekly usage. Plus users, in his framing, are relatively casual and newer, sometimes eating through a whole week’s usage accidentally then finding themselves confused.

The Second Reason Is Doing Some Work

The compute argument is honest and uninteresting. Load smoothing is a real operational problem and every metered service does it.

The second one deserves a closer look.

Describing Plus subscribers as relatively casual and new is accurate for the median user. It is completely wrong for the people this actually affects. A casual user does not exhaust a weekly Codex allowance. Somebody running long agent sessions does. That person is a developer paying twenty dollars a month for a tool they use professionally.

One user claimed they could burn an entire weekly Codex allowance in about an hour of heavy use, saying they had tested it. That is not carelessness. That is a quota sized for a different kind of customer.

The five-hour window does prevent the confusing experience Sottiaux describes, where somebody hits a wall on Wednesday with nothing left. It does that by making sure you hit a smaller wall more often instead.

Which is a reasonable design decision. It is also a rationing decision described in the language of user experience.

Pro Is Exempt. That Is the Point.

Pro at $100 and $200 keeps the five-hour gate switched off for the coming months.

The $100 tier arrived in April 2026 with roughly five times the Codex usage of Plus, access to exclusive models, plus unlimited Thinking mode. It now also carries the only escape from hourly restrictions.

So the segmentation is explicit. Casual users on one side with two overlapping limits, developers plus professionals on the other paying five times as much for uninterrupted agent time.

Nobody should be surprised by that. It is how every metered product eventually works. OpenAI has been unusually direct about saying so out loud rather than burying it.

The awkward part is timing. Plenty of people subscribed to Plus during the six weeks when the cap was gone, which changes what they were buying. Reporting on the reaction describes exactly that complaint appearing alongside the compute debate.

OpenAI Built Tooling to Import Claude Code Sessions

Here is the detail nobody covering this has connected.

Over the past few months OpenAI has been building native migration support into Codex for importing sessions from other agents. Pull requests in the Codex repository cover a codex-external-agent-migration crate, session detection with compaction handling plus token usage seeding, background import with deduplication, plus preserving titles from Claude Code session metadata. OpenAI also ships a curated migrate-to-codex skill with scan-only, plan, doctor, plus dry-run modes.

That is an engineering investment aimed squarely at making it painless to leave Claude Code for Codex.

Then they reinstated the limit that made Codex attractive in the first place.

Both decisions are individually defensible. Together they describe a company that spent months lowering the switching cost inward, then raised a friction some of those switchers moved to avoid.

This Has Been a Retention War for Six Weeks

Line the dates up and the pattern is obvious.

On July 12, OpenAI removes the five-hour cap for Plus. A week later Anthropic extends Claude Code weekly limits 50% higher through August 19. By July 29 Sottiaux is teasing reinstatement, then postponing it. Early August brings two more resets from OpenAI, with Codex crossing 15 million users on the 13th. The next day Anthropic ships an auto-continue checkbox in the Claude Code desktop app that resumes a stalled session automatically once your usage window resets.

That last one is worth pausing on. Anthropic’s answer to limit friction was not more quota. It was removing the manual step of restarting after a reset.

Then August 25, the cap comes back. September 1, Sonnet 5 gets more expensive.

Six weeks of quota generosity, timed resets, friction reduction from both sides, ending with both of them tightening in the same week. Our piece on Anthropic’s first profitable quarter covers why: private companies price for growth, public ones price for margins, plus both are weeks from public markets.

The generous period was not a policy. It was a phase.

What You Can Actually Do About It

Four options, in rough order of effort.

Do nothing and adjust your rhythm. Five hours is a long session. If you work in bursts with breaks, you may never notice. The people who will notice are the ones running multi-hour autonomous agent work.

Worth knowing what the window actually measures too. It is five hours of active use rather than five hours on the clock, so a session with long gaps while you read a diff or take a call stretches considerably further than five hours of wall time.

Move sessions between agents when you run out. This is the interesting one. The tool got published eight days ago.

session-migrate moves local coding-agent sessions between Claude Code, Codex, Pi, OpenCode, GitHub Copilot CLI, Antigravity CLI, plus Cursor Agent. MIT licensed, 40 stars, last pushed today, Python 3.11 or higher on Linux for now.

The command is short:

smigrate transfer SESSION_UUID --from claude --to codex --cwd "$PWD"
codex resume NEW_SESSION_UUID

It also inspects a transcript without printing the conversation, plus keeps a searchable catalog so you can find an old session by describing it rather than remembering a UUID.

Its author’s own framing is the case for it: if you started in Claude and ran out of usage, convert the session to another agent in one command.

The paid options

Upgrade to Pro. Five times the Codex usage, no five-hour gate for now. At $100 a month that is a real calculation rather than an obvious one. It depends entirely on how many hours you actually lose to limits.

Run the model locally. We covered an inference engine from Berkeley that runs a 35B model at 39.3 tokens per second on an 8GB laptop GPU, against a 33 token median in real Codex production traces. It exposes Anthropic-compatible and OpenAI-compatible APIs, so Claude Code points straight at it. No quota exists because no vendor is metering you.

That last option was theoretical a month ago. It is now a hardware purchase with a payback period.

What session-migrate Actually Does

Since it is the practical answer here, worth more than a mention.

The tool reads a coding agent’s native session file, converts it into another agent’s format, then writes it where that agent expects to find it. Your conversation, tool calls, plus working context move across intact. The receiving agent resumes as though the session had always been its own.

Seven agents are supported: Claude Code, Codex, Pi, OpenCode, GitHub Copilot CLI, Antigravity CLI, plus Cursor Agent.

There is an inspect command that reads a transcript and reports on it without printing the conversation, which matters if you are working on anything you would rather not dump into a terminal.

The catalog is the part that will get used most. It indexes sessions across your machine, then searches them by description rather than by UUID. Something like smigrate catalog search "authentication refactor" finds the session, then transfers it by catalog ID instead of a hex string you never wrote down.

Requirements are honest and narrow. Python 3.11 or higher, Linux only for now, installed through uv or pipx. The demo on the project page uses synthetic credential-free fixtures rather than a real session, which is the right call for something that handles conversation history.

Verified before writing: 40 stars, MIT licence, created August 17, last pushed today, two forks, zero open issues. Nine days old, which is worth knowing before you trust it with work in progress.

The Honest Caveat on All of This

One thing to hold alongside everything above.

Every number in this article is a policy that somebody can change. The five-hour limit was removed in July, restored in August, teased then postponed in between. Weekly allowances were reset early at least three times in six weeks. Anthropic extended limits 50% for exactly one month, then let that expire.

None of that is bad faith. It is a young product category where nobody knows what the right numbers are. Every vendor is adjusting in public.

But it means a decision made on quota terms is a decision made on sand. The developers who moved to Codex in July were reading the same public information everybody else was. They were not wrong at the time. The terms simply moved.

Which is the argument for the fourth option in the list above, more than any specific tool. A local model has no quota because there is no vendor. That trade costs money up front, hardware, plus setup time. What it buys is terms nobody can revise on a Tuesday morning.

The Comparison Everyone Is About to Redo

If you are choosing between subscriptions this week, the ground moved under both.

Codex on Plus now carries a five-hour window plus a weekly allowance, both active, with Pro at $100 as the escape hatch. Claude Code has an auto-continue feature that reduces the pain of hitting a wall without raising the ceiling, plus Sonnet 5 pricing rising on September 1.

Neither is clearly better. The honest answer depends on how you work. Somebody doing bursts of an hour or two barely notices a five-hour window. Somebody running overnight agent jobs hits it every time.

Our comparison of ChatGPT, Claude and Gemini covers the model differences, plus the Claude Pro review goes through what that subscription actually includes. Both are worth rereading against this week’s changes rather than last month’s.

One thing worth saying plainly. Any article comparing these two on limits has a shelf life measured in weeks. The last six weeks produced at least seven separate changes across both vendors. Treat every quota number you read, including the ones here, as a snapshot.

Why This Keeps Happening

Step back from the specific policy.

Agent workloads broke subscription pricing. A chatbot conversation costs a predictable, small amount. An agent session running for hours with tool calls, subagents, plus repeated context consumes orders of magnitude more. Consumption also varies enormously between users on the same plan.

Flat monthly pricing assumes rough parity between customers. Agents destroyed that assumption. One Plus subscriber might use a hundred times the compute of another for the same twenty dollars.

So vendors are all searching for the mechanism that keeps a simple price while limiting the tail. Weekly caps, rolling windows, credit purchases, tiered access, per-token routing. Nobody has landed on the right answer, which is why the policies keep changing.

DeepSeek gave away an entire agent runtime for free then raised model prices four days later. OpenAI cut Luna 80% while metering Codex more tightly. Anthropic reduced limit friction without raising limits.

Every one of those is the same company solving the same problem from a different angle: get people using agents heavily, without letting heavy agent use destroy the unit economics.

The eventual answer is probably usage-based pricing with a generous floor, since that is where every other metered infrastructure product landed. Cloud compute went through exactly this, as did bandwidth before it. What nobody wants to be is first, because a flat twenty dollars is enormously easier to sell than a meter running in the corner of the screen.

So the interim is what you are living through. Flat prices with hidden meters, adjusted quietly, several times a quarter.

/separator

The Part Worth Keeping

In July, OpenAI removed a limit and some developers moved over because of it.

Tomorrow the limit returns. The reason given is that Plus users are relatively casual, sometimes using their quota accidentally.

The developers who switched are not casual, nor did they use anything accidentally. They read a policy, made a decision, then the policy changed six weeks later.

None of which makes it the wrong call. Compute is finite, the plan is twenty dollars, so something has to give.

But if you picked a tool because of a limit, the lesson is not which vendor to pick next. It is that a quota is not a feature. It is a setting somebody can change on a Tuesday.

Check the terms again before September. Both of them move.

Charts and Blocks

Six weeks of policy changes

What applies to which plan