I Feel Vibe Coding Has Hit Its Ceiling

Over the past two weeks, the reputation of Claude Opus 4.7 has deteriorated sharply. Ever since OpenAI released GPT-5.5, public opinion has tilted almost entirely in one direction — both domestic and overseas communities are unanimously bashing Opus 4.7 while praising GPT-5.5.

The Reputation Crash of Opus 4.7

From what I have observed, there are two main reasons.

First, Opus 4.7 switched to a new tokenizer, which caused many previously well-tuned prompts to lose their former effectiveness. Second, the new tokenizer increased the overall token consumption for the same content, so users hit their quota limits faster and naturally feel that the experience has gotten worse.

The newly released GPT-5.5 goes the other way: higher output quality, faster response speeds, and crucially, far more available quota within the same-tier plan compared to Claude.

After GPT-5.5 went live, Anthropic has been scrambling to make up ground. First they published a lengthy post-mortem on the “Claude model dumbing down” issue, pointing to two causes: the default thinking depth was quietly lowered, and a bug was degrading output quality. Then last week, news broke that Anthropic struck a partnership with SpaceX, with the latter providing 220,000 GPUs and the matching power supply. In the same week, Anthropic announced that Claude’s 5-hour quota window would be doubled (though I personally did not feel an obvious doubling in practice), and today they further announced a 50% increase in the weekly limit.

This entire sequence of moves makes it clear just how much pressure GPT-5.5 has put on Anthropic.

The Shifting Balance Between Codex and Claude Code

Recently, OpenAI has also been cracking down hard on cheap shared accounts. Those Plus accounts that go for a few bucks a month basically don’t survive more than a day or two, so the actual cost of Codex is now roughly on par with Claude Code — both following the $20 / $100 / $200 three-tier pricing.

But there is a key difference: when paying through in-app purchases on Android or Apple stores, Codex maintains its original pricing, while Claude’s $100 / $200 plans incur an additional 25% “Apple tax / Google tax.” Doing the math, Claude Code’s actual cost becomes $125 / $250 — noticeably higher.

For developers without overseas bank cards, this is especially painful, because compliant signup almost only leaves in-app purchase as the only viable path.

Layering on a few more factors:

• Codex offers more generous quotas;
• Claude occasionally has account-banning incidents, while OpenAI is hardly ever heard of banning users who pay through legitimate channels;
• Codex has recently caught up on Computer Use, remote control, and similar capabilities — it is closing the feature gap with Claude very quickly.

All things considered, Codex has now surpassed Claude Code on cost, quota, and stability — and the gap on all three is shrinking fast.

Why I Feel Vibe Coding Has Hit Its Ceiling

With all that build-up out of the way, let’s get back to the title itself.

All the major AI vendors are raising prices. Domestic models once famous for their cost-effectiveness — GLM, Qwen, and most recently Doubao — have all started charging. For enterprises, bulk-purchasing AI is a significant expense: the cheaper models are barely good enough, while the better ones make costs balloon.

An awkward cycle is starting to take shape: companies assume that with AI they no longer need as many people, so they lay off staff and roll out AI tools at scale; the result is that the remaining employees take on heavier workloads, while the company’s total cost does not actually go down — it goes up, because the AI spend can even exceed the previous labor cost.

The more fundamental reason is that today’s AI models have not yet reached the stage of “iterating on a product entirely on their own without humans in the loop.” Over the past few weeks I have been seriously trying Vibe Coding, and below are several problems I personally ran into.

1. The Cost of Human Intervention Is Still High

As I said above, AI has not yet developed to the point of fully autonomous iteration. It can amplify individual productivity, but it is hard to truly “do two things in parallel” — like letting AI run a task while you go read documentation or watch a tutorial on the side.

I have tried this work mode, and the result is that AI constantly pops up asking you to make decisions or grant permissions. Every interruption drives up the intervention cost — you don’t actually learn anything, and you don’t get to finish a single video either.

2. Product Direction Starts to Drift

As productivity gets amplified, sometimes you can’t even think of what feature to build next. So you just let AI propose the requirements. The problem is that AI does implement them, but in the end you don’t fully know what features were added or what product direction was taken — you have already lost control at the product level.

3. Testing Cannot Be Fully Handed Off

Feature implementation and product planning can be handed off to AI to some extent, but testing currently cannot. Computer Use can indeed run some UX tests and fix some surface-level bugs, but it is nowhere near covering the full scope of testing — especially around business edge cases and exception paths.

4. Homogeneous Competition Arrives Too Fast

The boost in productivity brings a side effect: once one product goes viral, a flood of copycats appears almost immediately, at an unbelievable speed — within less than a day the market is full of homogeneous products. For independent developers, this environment is extremely unfriendly.

5. Corporate Scenarios Are Still Limited

Inside a company, Vibe Coding’s usable scenarios are similarly limited. Letting AI complete version iterations based on a single product prototype is still unrealistic — most of the key decisions still have to be made by humans.

6. Project Migrations Still Require Deep Involvement

I once did a not-so-large project migration: from Vue 2 to Vue 3. Because Vue 3 contains breaking changes, almost every file needed to be rewritten.

The project had a little over seventy pages. I had AI carry out the migration and wrote a lot of prompts, but the result was that not a single attempt was able to complete the migration in a single conversation window — it had to be broken down step by step.

This was quite disappointing. The scenario I had originally hoped for was: write a prompt, leave it running, go do something else, and the AI just migrates Vue 2 to Vue 3 on its own. Reality clearly isn’t there yet.

Conclusion: AI Is a Lever, Not a Replacement

After this round of hands-on testing, I have a much clearer view of AI’s current positioning: it is an assistant, not a replacement.

The more skillfully someone uses AI, the bigger their opportunity. But pinning your hopes on “AI running the business on its own” — at least at this point in mid-2026 — does not hold up.

What is more worth paying attention to is its nature as a “cognitive lever”: using it to learn and to expand your knowledge boundaries gives you an efficiency gain measured in orders of magnitude. This may well be the safest and most compounding way to use AI right now.