Is the Era of Cheap AI Over?
A $500 ChatGPT plan on one side, Anthropic cutting Opus prices on the other. Which direction is the future?
Introduction
About three years ago, the big question was: should I pay $20 a month for ChatGPT Plus? Today the question is completely different. In the last week of September 2026, leaks surfaced about a new OpenAI plan called ChatGPT Pro Max, priced at $500 a month. Almost the same week, Anthropic launched Claude Opus 5.5, at a lower price than the model it replaced.
Two stories pointing in opposite directions. One says AI is becoming a luxury product; the other says it keeps getting cheaper. In this article I try to make sense of what is actually happening, and what it means for us as developers and users.
What is ChatGPT Pro Max?
On September 24, 2026, leak trackers (TestingCatalog and researcher Tibor Blaho) spotted strings in ChatGPT’s code pointing to a new plan called Pro Max, priced at $500 a month, or $6,000 a year.
To put that number in context:
- Plus: $20 a month.
- Pro: originally $200, later joined by a $100 tier.
- Pro Max: $500, which is 25 times the price of Plus.
The interesting part is that the main selling point isn’t a smarter model, it’s speed: the fastest experience in Codex (OpenAI’s coding agent) and Work (its tool for long research and writing tasks). A few days later, OpenAI merged a change into its public Codex repository adding support for a new plan type shown as “Pro (Max)” and renaming the existing tiers, which is the strongest evidence so far that the plan is real.
Note: At the time of writing, the plan is based on leaks and code changes, and it does not yet appear on OpenAI’s official pricing page. Details may change at the official launch.
Why is OpenAI selling “speed”?
To understand this plan, we need to look at how the way we use AI has changed.
In 2023, usage was mostly question and answer: you send a message, the model replies in seconds. Today the heaviest usage comes from agents: you give Codex a task like “fix this bug and write the tests,” and it works for minutes or even hours, reading dozens of files, running commands, and trying again.
That kind of work burns an enormous amount of compute. And when an agent runs for an hour, getting the same result in 20 minutes has real value for a professional developer or a company. So OpenAI is betting that some people will happily pay $500 to save time, the same way people pay extra for express shipping.
There is also speculation that the plan could run on Cerebras chips, which specialize in fast inference, but that remains an unconfirmed guess.
The opposite direction: Anthropic and Claude Opus 5.5
On September 22, 2026, Anthropic released Claude Opus 5.5, and the headline message was clear: better performance at a lower cost.
Prices per million tokens:
Input
Opus 5 :$5
Opus 5.5 :$4
Output
Opus 5 :$25
Opus 5.5 :$20
Get Khalid’s stories in your inbox
Join Medium for free to get updates from this writer.
Cache reads
Opus 5 :$0.50
Opus 5.5 :$0.20
Anthropic says Opus 5.5 costs about 40% less to run than Opus 5 on typical workloads. That figure doesn’t come from the price cut alone (which is 20%), but also from the model needing fewer tokens and fewer steps to finish the same task. The company also says it generates output more than 30% faster.
Even more striking: Opus 5.5 matches or beats Claude Fable 5.1, Anthropic’s top model, on most published benchmarks, while Fable 5.1 costs $10 for input and $50 for output. In other words, Anthropic shipped a much cheaper model that competes with its own most expensive one.
Notice the big cut in cache read pricing too (60%). It looks like a small detail, but it matters a lot for agents and coding tools, because they resend the same context (code, instructions, conversation history) over and over. In that kind of work, cache reads make up most of the bill.
Put simply: while OpenAI sells speed to those willing to pay more, Anthropic is trying to make each task cheaper for everyone.
So, is cheap AI over?
My answer: no, but the meaning of “cheap” has changed.
We are actually living through two trends at the same time:
1. The price of intelligence itself is falling. Every new generation of models delivers better performance at a lower price per token. What cost $50 a few months ago can now be had for something close to $20. And at the other end, OpenAI itself launched very cheap plans like ChatGPT Go, priced at a few dollars in markets like India.
2. But how much we consume is growing even faster. We no longer send one question; we run agents that work for hours and burn through millions of tokens. So even as the per-token price falls, the total bill for a heavy user stays high.
The result is that the market has split into tiers:
- A free or very cheap tier for most people: questions, writing, summaries.
- A middle tier ($20 to $200) for professionals.
- A premium tier ($500 and up) for people who value time over money, like companies and developers building entire products with agents.
In other words, the ceiling is rising and the floor is dropping. Everyday AI is getting cheaper, while “unlimited” AI is getting more expensive.
What does this mean for you as a developer?
If you’re building a product on top of AI, or using these tools every day, here are some practical lessons:
- Measure cost per task, not price per token. A model that costs more per token can end up cheaper if it finishes the job in fewer steps.
- Use caching. If your app sends the same instructions or context repeatedly, prompt caching can cut your bill significantly.
- Use batch processing for work that doesn’t need an instant response. For example, Opus 5.5 in batch mode costs half the standard price.
- Pick the model for the task. You don’t need the most powerful model to classify a message or summarize a short text.
- Only pay for speed when it’s actually worth the money. A $500 plan makes sense for a company where saved time is worth thousands, but it rarely makes sense for a student or an indie developer just starting out.
Conclusion
ChatGPT Pro Max isn’t the end of cheap AI; it’s the start of a phase where time and speed are priced as a product of their own. Meanwhile, Claude Opus 5.5 shows that the real competition may be about efficiency: who gives you the same intelligence for less.
For most of us, that’s good news: tools that were reserved for big companies a year ago are now within everyone’s reach. The question worth asking isn’t “is AI expensive?” but “what is my time worth?”
What do you think? Would you pay more for speed, or do you prefer the cheaper, more efficient model? Let me know in the comments.
