OpenAI Merged ChatGPT and Codex Into One Allowance

OpenAI sent subscribers an email last week with one sentence that matters more than the rest: "Your use of ChatGPT and Codex will draw from the same allowance."

One barrel. Two taps. Whatever you pour out through chat, you don't have for Codex, and the reverse.

I've been watching the shared usage allowance story for a few weeks because I live inside these limits every day. And the pattern is clearer now than any pricing page makes it.

What Actually Changed

Here's what I can verify. OpenAI's own pricing page for Codex now says ChatGPT Work and Codex share usage. An email to subscribers, reported by BleepingComputer on November 1, says the same thing. Unlimited Instant is excluded.

Around the same time, OpenAI doubled the Work and Codex allowances on its Pro plans from October 29. That came after backlash over a planned cut of roughly 50%, which they reversed.

What I could not verify: the exact start date for the merge. One outlet says December for Plus, Pro and Business. The help article wasn't readable for the person who reported it. Treat the date as unconfirmed.

Earlier this year the Plus plan went through its own whiplash. The five-hour window was removed in July, then brought back on August 25.

So in about four months: limit removed, limit restored, cut announced, cut reversed, allowance merged. That's not a pricing page. That's a dial someone is turning live.

The Quota Is the Product

Everyone argues about which model is better. Meanwhile the thing you actually buy is a number of tokens per window, and the vendor can change that number whenever they want.

I pay for Claude Max and I've felt this from the other side. I wrote about it when Anthropic split programmatic usage into its own credit pool in Anthropic Put Agent Calls on a Meter. Different company, same move: make the invisible limit visible, then reshape it.

The model is a commodity that gets cheaper every quarter. The allowance is the lever.

Why Merging Is Smart for Them

A merged pool does three things for OpenAI.

First, it smooths demand. Chat users are bursty. Codex users run long, hungry sessions. Put them in one pool and the average heavy user is subsidized by the casual one who barely opens the app.

Second, it makes the coding tool feel like an upgrade to the plan you already have, not a second bill. That's good for adoption.

Third, it makes your usage harder to reason about. One number goes down faster, and you can't easily tell which tap did it.

That third one is not a conspiracy. It's just what happens when two products share a meter.

The Uncomfortable Take

Most builders see this as a downgrade story. I think the real problem is different: you are building a workflow on top of a limit you don't control and can't audit.

If your agent loop, your nightly job, or your client demo depends on a subscription allowance, you have a hidden dependency. It has no changelog, no SLA, and no version pin. It can be doubled on a Thursday and cut on a Monday.

The people calling this a rug pull and the people calling it a gift are making the same mistake. They are both treating a consumer plan as infrastructure.

I'd tell you what I tell myself. Use the subscription for the work where a human is watching. Move anything unattended to metered API keys, where the price is written down and the failure is a bill, not a silent throttle.

What I'm Doing About It

Three changes, none of them dramatic.

I log tokens per task. Not for a dashboard, just a plain JSONL file per run, so I know what a unit of work costs. If you run agents, a flight recorder makes this trivial.

I keep one fallback model wired in for every automated flow in n8n. If one vendor tightens a limit, the workflow switches instead of stalling at 2am.

And I stopped planning around "unlimited". Nothing is unlimited. The word just means the limit hasn't been explained yet.

# one line per task: what did this actually cost?
echo '{"task":"nightly-triage","tokens_in":18240,"tokens_out":2310}' >> runs.jsonl

Who Wins From a Shared Pool

Light users win. If you mostly chat and occasionally run Codex, a bigger shared allowance probably works in your favor.

Heavy builders lose predictability. When two products drain the same barrel, a long coding session can eat the capacity you needed for something else, and you find out when it stops.

The honest read is that nobody outside OpenAI knows the exact numbers per tier, and OpenAI isn't publishing them. Anyone telling you precisely how much you lose is guessing.

Meter your own work. Keep an API key in reserve. Never build on a limit you can't read.