Something that was both my prediction and my concern a few months ago is materializing. I was holding out hopes I might be wrong, but the probability of that is growing lesser by the day.
For months I've said that the age of AI subsidy will come to an end. Today, OpenAI made that fist move. But we can expect other labs to follow suit really soon.
Here's the email I received today (see screenshot) from OpenAI that proves it.
They even do the whole sandwich method too.
They started by talking about GPT-6.1 Sol and how capable it is at how low the cost (and to be fair, it IS much lower in cost)
Delivered the blow by saying they're now halving the usage limit for all $200/mo Pro users (which I have four of)
Gave "62,500 usage credits to ease transition"
And briefly mentioned "Dots" as if its something new - except its simply a change in UI, modelling after the success of Grok Bot and Hermes
In return, they are also introducing a new $500 per month subscription with only 25x usage
The notice is 24 hours.
We used to have 20x usage at $200 per month.
Now we have 10x usage at $200 per month, and have to pay $500 per month if we want the 25x.
It should also be noted that just last week, OpenAI quietly informed all their users via the ChatGPT app that their Terra models will be retired soon.
This announcement from OpenAI is from my perspective a move they make to regain market leadership.
In the same way Anthropic did with Project Glasswing, and SpaceXAI with Grok Bot, OpenAI is setting a new precedence that other labs will model after really soon.
I am more concerned than ever.
What's wrong with simply NOT using frontier models?
Nothing wrong, but I am observing a pattern here.
GPT 5.4 was more reliable for me at staying focused and not drifting across days of work - until GPT 5.5 came out.
When GPT 5.6 came out, all the previous models became suddenly and noticeably dumber.
(and you need to forget about benchmarks because benchmarks measures models at their "best", not at their day-to-day)
All of a sudden, they can't get work done anymore.
They would drift, forget, make me repeat myself, and make mistakes that we need to reverse.
A 21 day shipment crisis I experienced happened in July because of this.
More recently, in an attempt to get move away from the frontier models, I tried using the open weight models like Deepseek V4.1, GLM 5.3, and Kimi K3.
All of these are supposedly "frontier level."
Yes - in benchmarks.
But again - benchmarks.
In real life application, I find myself having to babysit them for hours and hours just to prevent them from making mistakes.
That's untenable.
The difference between frontier models and open weight models are night and day, and the labs know that.
In my socials, I am beginning to see people running up 8x $200/month.
I am running 4x $200/month usage and I thought I'm a beast...
Our options are getting cut, most of us are simply and constantly reacting to what the big labs do, and they know it.
And they are doing everything in their power to make sure they stay in power - including attempts to block competitors by influencing government policies.
The gap between frontier and open weight is STILL closing...
But if I'm being honest, its starting to look like the gap isn't closing fast enough.
At the same time I can't help but feel that the pace of frontier AI is accelerating so fast that "gap closing" might just be completely irrelevant because for every 1 month they catch up the frontier is putting in 1 week extra in distance.
I hope I'm wrong on this too, and I hope open weights continue to improve.
Anthropic is still dealing with the whiplash of how poorly Opus 5 performed, and while they are regaining the trust of the market with Opus 5.5, they won't do much to hurt that rapport.
But eventually, they will also follow the precedence. They have all the incentives to do so.
Whatever project you are building right now, whatever pace we think we are being fast at, the big labs are pacing it even faster.
And yes, I think that the recent Dario Amodei article on Pacing The Frontier is total bullshit - I don't believe for a second they actually intend to pace the frontier.
Proof:
One week all the leaders agreed on pacing the development of AI
The following week all the big labs released new models
This week most of the big labs released yet another new model
Pacing, my ass.
So whatever project you are building right now, it's time to pick up the pace.
AI is getting expensive.
FELIX