(Looking at this official comparison chart, it’s obvious that GPT-5.4 was never positioned as a free-to-play toy from the start)
March in Shanghai rarely sees such a sunny day, and the 14°C weather is perfect for a stroll. But I reckon many developers glued to their screens this morning aren’t in such a bright mood.
Just a few hours ago, OpenAI pulled the plug on free users without any warning—to be precise, they revoked the calling quotas for GPT-5.3 and GPT-5.4 models on free Codex accounts.
Just How Expensive is a Free Lunch?
I had just poured coffee beans into the grinder this morning when my phone started buzzing non-stop. Several indie developer friends were venting in group chats, complaining that their overnight refactoring tasks threw a quota exceeded error, halting right in the middle.
It’s certainly frustrating. After all, OpenAI just recently made a huge splash announcing the Codex app, specifically granting limited-time free trial quotas to ChatGPT Free and Go users. Everyone was having a blast using GPT-5.4’s native Computer Use and Tool Search features. Some even tossed migration work that usually takes a week to these agents for concurrent processing.
Who would have thought this honeymoon period would be as short as a weekend. But if you think about it, there were early signs of this across-the-board cut. Frankly, it comes down to the oldest excuse in the book: it’s simply too expensive.
Doing the Simple Math
Let’s look at this from another angle: do you really think a commercial company can keep footing the electricity bill for developers worldwide forever?
Just look at the actual pricing of these two models and it all makes sense. GPT-5.4 costs $2.50 per million input tokens, and a staggering $15 per million output tokens. Although GPT-5.3-Codex is slightly cheaper on the input side, it’s still nothing to scoff at. However, the real killer isn’t the unit price, but how Codex operates now.
In the past, using AI to write code meant you asked a question, it gave a snippet—a Q&A might consume a few thousand tokens at most. But today’s Codex is no longer a simple auto-completion tool; it’s a multi-agent environment that uses code as a tool. You give it a requirement, and it automatically searches docs, runs tests, and if it spots an error, it loops back to fix it. Once this cycle kicks off, token consumption skyrockets exponentially. The machine frantically self-correcting in the cloud is practically a tireless money-burning furnace for OpenAI’s servers.
Trading compute for testing and interaction data from free users, then decisively cutting the free channel once enough data is collected or the server bills get out of hand—it’s a very pragmatic playbook.
Can Switching Tools Solve It?
Someone is bound to say: since OpenAI is no longer allowing freebies, why don’t we just switch to other tools?
To be blunt, the current alternatives aren’t exactly worry-free either.
A few days ago, to test a new framework, I deliberately switched to a few other mainstream competitor models to run a slightly complex frontend architecture migration. For instance, a star model known for long contexts was undeniably fast, but by the time it hit the third file association, the context started suffering from mild hallucinations, outputting API parameters from a previous generation. The latest model from a top-tier Chinese tech giant held its own in pure algorithmic logic, but once it required linking up with a browser for UI testing, its native capabilities fell noticeably short.
GPT-5.4’s massive 1.05M context window truly offers a frustratingly reliable stability when handling the dirty work of large codebase analysis. Leaving this ecosystem means you either endure higher communication costs or manually patch up the edge cases the AI misses.
A Few Things I Sometimes Ponder
(Sighs helplessly) I sometimes wonder if this is a clear signal: the romantic illusion that “one person with a free AI can outcompete a whole team” might be coming to an end this spring.
As software development truly enters the automation era, the cost of cognition is being substantially converted into the cost of compute. In the past, large corporations built their edge by stacking manpower; in the future, it might be about crushing the competition with API quotas. If the token cost of writing a complex piece of business logic eventually rivals the cost of outsourcing to a junior developer, won’t the innovative capacity of indie developers and small teams be strangled by these exorbitant bills?
Perhaps I’m overthinking it. Maybe in six months, hardware and inference architectures will go through another round of optimization, or an incredibly capable small-parameter model will emerge from the open-source community to completely flip this expensive banquet table over.
Having said all this, I don’t really have a definitive conclusion ¯\(ツ)/¯.
My coffee has gone completely cold, I need to brew another cup. As for how to push forward with tomorrow’s work, everyone should honestly just check their remaining credit card balance in the backend.
References:
- GPT-5.4 vs GPT-5.3 Codex Coding Capability Comparison: 6 Benchmarks Reveal the Winner
- Introducing GPT-5.3-Codex – OpenAI
- Introducing GPT-5.4 – OpenAI
- Codex Pricing – OpenAI for developers
—— Lyra Celest @ Turbulence τ.
