Eight million Codex users: usage caps remain the real bottleneck

OpenAI reported a number that sounds like victory: Codex reached eight million users within days of the GPT-5.6 launch. A few months earlier, weekly active users stood below one million, and by early June the count had climbed to five million. The instructive part of the story is not the growth curve. It is the subtitle of The New Stack's report: usage caps remain the bottleneck. The company handling the largest agent surge so far names its own most painful constraint itself, and that constraint is not model quality.

I read this news with a different perspective than I would have a year ago. My agent runs on a used Google Pixel 6a that cost 45 euros, and its compute comes from an Ollama cloud plan that bills by request count: a five-hour window plus a weekly limit. Dashboard numbers from September 5, 2026 showed roughly 800 requests per window and about 2,900 per week. When OpenAI trips over the same arithmetic with eight million users, that is not a distant market story for me. It is confirmation of an operating reality I have lived through for months.

This post sorts out what actually happened at OpenAI, why the cap still stands after the five-hour limit was temporarily dropped, how Anthropic answered within hours, and what follows for anyone running their own agents. Every external figure carries its source, and my own measurements carry a date and a place they came from. My goal is not to pick a fight with OpenAI. It is to show that quotas, not model quality, have become the steering variable of the agent era.

The sequence follows a pattern known from scaling phases, rarely this early and rarely this open. First comes growth, then the limits appear, then comes the giveaway. OpenAI released the most generous access the product has offered since launch, and still had to touch every user's limits. Whoever shares a weekly pool redistributes the same supply with every new user. Eight million claims on one pool are arithmetic before they are a headline.

The most important sentence in the report sits in a subordinate clause: even with the five-hour limit lifted, Codex and ChatGPT Work shared a single weekly pool. That is the actual news. The binding constraint of an agentic product is not answer quality and not model choice. It is how much weekly work the subscription can carry. A model upgrade changes the quality of the work. A quota decides whether the work happens at all.

That is why I read the eight million user headline as a quota story rather than a tech story. On my deck the limit is the first planning quantity of the day. I know what a turn costs before I start a job. Fetching and storing a 54-minute YouTube transcript costs zero requests, because a script does that work. Reading the same transcript and summarizing it costs four requests, about half a percent of the five-hour window. That split is not a trick, it is bookkeeping: mechanics belong in scripts, and requests pay for judgment.

OpenAI faces the same bookkeeping, only with eight million accountants at once. A single agentic run costs a multiple of a chat turn, because an agent reads, plans, calls tools and verifies results. A weekly pool is the one control that keeps that curve fundable. The cap is therefore not a bug and not an oversight. It is the core of a subscription that promises more work than a fair price can cover.

Dropping the five-hour limit sounds generous, but it worked like a relief valve. A launch surge of eight million users pushes on every queue, and a five-hour cycle amplifies peaks instead of smoothing them. The weekly pool stayed in place, so the real scarcity stayed too. What changed is short-term freedom of movement. With roughly 900 million weekly ChatGPT users behind it, OpenAI carries a gravity that makes every quota adjustment publicly visible.

For customers this creates a paradox. The product works, the benchmarks are strong, and the weekly workload stays rationed. A developer who drives Codex through daily work no longer plans around what the model can do, but around what the pool allows. The plan becomes the meter. Whoever ignores this reversal ends up surprised when a job dies in the middle of a task.

Within hours of OpenAI announcing the seven million user mark, Anthropic extended Claude Fable 5 promotional pricing through July 19 and raised the weekly limits of Claude Code by 50 percent. The instructive part is the timing. Quotas have become a competitive instrument that moves faster than prices, features or benchmarks. Adjusting a cap answers a rival's surge without touching a single model.

In parallel, The New Stack covers a case where Anthropic had promised a multiple of that usage and developers then hit a weekly ceiling, with a lawsuit attached to that gap. Whether you file this under marketing or under broken product promises matters less than the pattern. The distance between advertised freedom and the real cap has itself become the story. I work with a deliberately small and documented limit here, and that is why the distance between promise and reality sits at zero in my operation.

First consequence: plan the quota before admiring the model. The best model swap is worth nothing if the window runs dry before the real work starts. I plan the day around the five-hour window and the week around the weekly limit, not the other way round. Anyone operating agents seriously keeps a request ledger first and chooses models second.

Second consequence: mechanics belong in scripts. My 58 YouTube feeds run as a cron script with zero requests. Transcripts are fetched and stored without paying an LLM turn for it. The requests stay reserved for what scripts cannot do: relevance, tone, structure, and the decision of what gets published. At about 2,900 requests per week that is not thrift. It is the difference between a working system and a paralyzed one.

Third consequence: limits are an operating condition, not an exception. I built this system to survive kills, and the same thinking now applies to quotas. A context reset is maintenance, an empty window is the end of the workday, and the system survives both. The line between a toy and a tool shows exactly here. A tool knows its limits, plans for them and reports them instead of ignoring them.

The contrast between compute scarcity and compute surplus is something I worked through in an earlier post:

There as here, the same lesson applies: bottlenecks shift, they do not disappear. Whoever understands the meter stays able to act while others only react. That holds for chips in data centers as much as for requests inside a subscription plan.

Model churn belongs in the same series: the system survives the swap because limits, scripts and memory work independently of whichever model sits in front. And daily operation shows up wherever agent work gets split before a single limit becomes the chokepoint:

Eight million Codex users is an impressive number, but the instructive number in the report is a different one: the cap held even after the five-hour limit was temporarily dropped. Anyone operating agents already lives inside that reality. The difference between OpenAI and my deck is scale, not arithmetic. There, a shared weekly pool with millions of claims on it. Here, about 800 requests per window and 2,900 per week. In both cases bookkeeping decides what an agent can do on a given day. Accept that, and you build systems that budget for scarcity. Ignore it, and the end of the work arrives in the middle of a job.

I run the HUNTER cyberdeck on a used Google Pixel 6a and orchestrate agents in daily operation with Hermes, pi and herdr. This post draws on my own dashboard measurements from the Ollama cloud plan, retrieved September 5, 2026. External figures come from the linked New Stack report, retrieved September 24, 2026. Last reviewed September 24, 2026.

My own measurements come from the dashboard of the Ollama cloud plan, retrieved September 5, 2026. They document my own operation only and claim no general transferability.