With OpenAI’s Codex, a $200 subscription now gets you roughly half of what it did a month ago, measured in API prices. OpenAI claims you’ll get more done than before because the models got cheaper. The community doesn’t buy it, and in the replies people mostly compare it to Claude Opus 5.5.
What’s changing
Thibault Sottiaux (known as Tibo on X), who leads the Codex coding agent at OpenAI, announced on September 29 that starting tomorrow the company is reopening the $200 Pro plan (roughly 4,000 Czech crowns) to new subscribers. New sign-ups and upgrades to this plan had been paused since September 10. Along with the reopening, however, the way usage is counted is changing.
Codex bills work in credits tied to the list prices of the API. Each model therefore eats into your subscription as much as it would cost if you paid per token. From now on, the $200 plan will contain work worth half as many API dollars as before. Tibo says so himself in the post, though in a phrasing that commenters immediately seized on: “when you do the math, it comes out to half.”
The tweet rests on OpenAI’s main arguments:
- The five-hour limit is not coming back. Until now, in addition to the weekly volume, you also had a cap on a five-hour window. That stays permanently abolished, so you can burn through your weekly volume whenever you like, say in a single long afternoon.
- Models are getting cheaper, and the savings are passed on to the subscription. Last week (September 22), OpenAI released the GPT-6 Sol and GPT-6 Luna models at half the price of the previous GPT-5.6 generation. Sol is the mid-tier model for everyday coding, Luna is the cheapest and fastest variant for small tasks, and with the Max version it’s very attractive for development; I happily use it myself.
- The company doesn’t want to artificially inflate API list prices to make the subscription look like a better deal. Over the long term, it expects the API to get cheap enough that the gap between the subscription and pay-as-you-go nearly disappears.
- The subscription will gain features that don’t count toward usage. Tibo didn’t say which ones. OpenAI’s DevDay, its developer conference, is tomorrow.
| Model | Input (USD / 1M tokens) | Output (USD / 1M tokens) |
|---|---|---|
| GPT-5.6 Sol | 4.00 | 20.00 |
| GPT-6 Sol | 2.00 | 10.00 |
| GPT-5.6 Luna | 0.20 | 1.20 |
| GPT-6 Luna | 0.10 | 0.50 |
How did OpenAI do the math for us?
OpenAI’s logic goes like this: you get half the dollars, but tokens cost half as much, so it comes out the same, and since GPT-6 Sol is better than GPT-5.6 Sol, you get more done. For Sol and Luna, that roughly checks out.
The catch is the flagship model, GPT-6 Astra. It didn’t get cheaper. Anyone who works mainly with it because they need the highest quality will, from tomorrow, effectively have half the volume of work. And Astra is exactly what programmers in the replies point to most often.
The second ambiguity: does the change affect only new subscribers, or existing ones too? Tibo writes about opening the plan to “new subscribers” and about the calculation “compared to the old plan.” Throughout the whole thread, he didn’t answer direct questions about whether existing users will stay on their original terms. I wouldn’t count on it.
Tommy Long noticed a third thing: if the “20x” plan (twenty times the limit of the Plus plan) is cut in half, it will be only twice as generous as the “5x” plan at $100. Double the price for double the volume.
How the community reacted
The thread has over 2,600 likes and almost 900 replies. The tone of the replies is mostly negative, and three motifs keep recurring.
Comparison with Anthropic. The most-liked reply, from user Chandraprakash, reads roughly “beat Opus 5.5 on performance per price and you won’t have to write long posts explaining yourselves.” Tibo answered with a single word: “Agreed.” That’s an unusually candid admission from a product lead. Other users write that they’re canceling their Codex subscriptions and switching to the Claude Max plan, because in their view Opus 5.5 is catching up to Astra in quality and offers more usage for the same money. Who knows, maybe these are the same users who were telling Anthropic the same thing a month ago…
Distrust of the “more work for fewer dollars” promise. Marcos Cámara summed it up with the image of a pizza that has four slices instead of eight, accompanied by a five-paragraph letter explaining why that’s better. Others point out that the whole argument rests on a promise of future price cuts, that is, on a bet about how quickly models will improve.
Suspicion of a push toward a pricier plan. Several people mention a $500 plan and read the change as nudging users toward it. Whether such a plan will really arrive isn’t known yet; OpenAI hasn’t officially announced it. But DevDay is on 9/29 (sic: tomorrow), and its unveiling is expected there.
There was also criticism of the timing. Opus 5.5 came out on the same day as GPT-6 Sol and Luna, and developers had been waiting for weeks for DevDay, hoping for cheaper usage. Instead, they got news of halved limits one day ahead.
What to take away from this
You have to give Tibo credit: he said it straight out and upfront, without hiding it in a footnote. His point (c) also makes economic sense: tying the subscription to the API price list and not inflating it is more honest than playing games with huge “virtual” discounts.
But subscribers don’t care about economic elegance. They care how many hours of agent work they get for $200. And here OpenAI is losing at the worst possible moment, when the competition offers a comparable model with looser limits. A programmer on the $200 plan is exactly the kind of customer who will run the comparison and switch within a single afternoon.
If you use Codex, I recommend waiting for tomorrow’s DevDay and only then deciding. Tomorrow we’ll find out what those “outside usage” features are and whether the change will also touch existing plans. If you already work mainly with Astra, be prepared for your weekly volume to last about half as long.
Sources: Thibault Sottiaux’s post on X and the replies in the thread (as of September 29, 2026), GPT-6 Sol and Luna pricing according to The New Stack. Plan terms change week to week; after DevDay on September 30, they need to be verified again.