GPT-6 Astra: OpenAI’s smartest model, half the messages
The model OpenAI itself calls its smartest ever landed in ChatGPT this weekend, and it will not cost you a cent extra. That sounds like a gift, until you check the meter: on every plan, GPT-6 Astra gives you roughly half as many messages as its predecessor. The subscription fee stays put while what you get for it melts away. The limit is the new price, and this is already the second time in a single week that a major AI player has pulled that trick.
What OpenAI actually announced
On 3 September, OpenAI published the GPT-6 Astra announcement, calling it “our most intelligent and aligned model yet”. Aligned, in this context, roughly means trained to do what you actually intend without going off the rails. OpenAI claims top-tier performance in computer use, coding, cybersecurity and science. The cybersecurity claim is not empty marketing: Astra is the first OpenAI model to reach the “Critical” level under the Preparedness Framework, the company’s internal risk framework for tracking how dangerous a model’s capabilities could become. No earlier model from the company had reached that level.
The rollout in ChatGPT began on 5 September, The Decoder reports, based on OpenAI’s announcement on X and the official pricing documentation: the Pro, Enterprise and Business Premium plans go first, with Plus and Business following a few days later.
The limit is the new price
That same pricing documentation contains the number that really matters. On every plan, the standard version of Astra allows roughly half as many messages per five-hour window as GPT-5.6 Sol:
| Plan | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|
| Plus | 10 to 100 messages per 5 hours | 5 to 45 |
| Pro 20x (heaviest Pro tier) | 200 to 2,000 per 5 hours | 100 to 900 |
The separate GPT-6 Pro variant is scarcer still and is metered weekly: 200 messages on the $200 Pro plan, 50 on the $100 one.
So your bill does not rise by a single dollar, and yet you pay roughly double per available message. That is the sly part of this price increase: it never shows up on any statement. The amount leaving your account each month stays exactly the same.
If you have followed any AI news this week, that will sound familiar. At Anthropic, Claude Code’s weekly limits drop a net 17 percent on 14 September while the subscription fee stays flat. The real news about Claude Fable 5.1 was the price there too, we wrote at the time. Two market leaders, one week, the same move: leave the price tag alone, tighten the meter. That is no longer a coincidence, that is a pattern.
The honest counterargument
There is a case to be made for OpenAI. A heavier model devours more compute: the servers do more work per answer, so the same machines can handle fewer Astra messages than Sol messages. Keep the limits where they were and you have to raise the price or make everyone queue. And there is a second mitigating factor: if Astra really is that much smarter, you may need fewer messages to get the same job done. A problem that took Sol three attempts might take Astra a single turn. If you mostly ask hard questions, you could even come out ahead on balance.
That “if” is carrying a lot of weight, though. How smart Astra actually is remains hard to verify independently today: benchmark firm Artificial Analysis overhauled its entire Intelligence Index on 5 September after the Astra scores raised doubts. Incidentally, after that revision Astra climbed to second place; the point is not that the model disappoints, but that even the measuring stick is moving. And the “smarter model, so fewer messages” arithmetic mainly holds for heavy tasks. If you mostly ask ordinary questions, you burn through messages at the same rate as before; you just have half as many of them.
What to make of all this depends on how you use ChatGPT. If you never hit your limit on Plus, little changes and you get a stronger model at no extra cost. If you regularly bumped against your ceiling, start counting over the coming days, because that ceiling now sits a good deal lower. Whether GPT-5.6 Sol will remain available as a more economical alternative is unclear for now; that too will determine how hard this cut bites.
The subscription price is quietly becoming the least interesting number on an AI pricing page. If you want to know what you are really paying, look at the meter: how many messages, per how many hours, from which model. That is where pricing happens now, neatly out of sight of your invoice.
