What Goes Around Comes Around: A New Model Every Month and a Half

The long-awaited OpenAI "potato" dropped today.

Turns out it's not GPT-6, but GPT-5.5.

I tried it out right after it came out and found that it fixed the thing that annoyed me most about 5.4: it finally started talking like a human.

Back in the 5.4 days, reading through Codex's output process was torture. Words piled on top of each other with no logic, status reports and execution notes jumping back and forth—you mostly couldn't tell what it was actually doing. 5.5 fixed this, and the readability of the entire workflow shot up immediately.

And then it gets funny.

Just a while ago I was telling my team: GPT 5.4 is stronger than Claude on certain complex tasks, but it doesn't talk like a human, so it's exhausting to use; Claude speaks human, it can guide you through, so we should still push Claude Code. Some colleagues went and canceled Codex and switched to Claude after hearing that. Today I have to send another message in the group chat: maybe just get both accounts and use them together.

Can everyone grasp how absurd this is?

A New Generation Every Month and a Half

5.4 came out in early March. I judged at the time that this wasn't a minor version—it was leaps and bounds stronger than the previous generation, Codex 5.3. Now it's April 24, and a new generation dropped after just a month and a half.

Anthropic's pace is even more extreme. Opus 4.6 came out in February, Opus 4.7 last week—two months in between. Even more absurd, on April 8 they officially released a model called Mythos Preview. The company says this is their strongest model to date: 93.9% on SWE-bench, and 31 percentage points higher than Opus 4.6 on the USAMO math olympiad.

But this model was not opened to the public. It was only made available through a program called Project Glasswing, with restricted access for over 50 institutions including Microsoft, Google, AWS, JPMorgan, Nvidia, and Cisco. Bundled with $100 million in usage credits.

The reason wasn't that they were keeping it under wraps. The company stated that Mythos Preview's vulnerability-hunting capabilities were absurdly strong—it could find what they described as "tens of thousands" of zero-day vulnerabilities in mainstream operating systems and browsers that ordinary bug hunters couldn't uncover at all. Fully releasing it posed too much risk. Last week's Opus 4.7 is essentially the "publicly releasable" neutered version of that same generation.

This is the rhythm now: a new generation is guaranteed every month and a half to two months, and top-tier versions are already being restricted from distribution under the rationale of "too powerful to release publicly."

Last Year It Was Three Months Minimum

This pace would have been unimaginable a year ago.

After DeepSeek R1 came out in early 2025, people discussed it for a solid six months. Talking about R1 in February, still talking about R1 in May and June. V3.1 in between was a minor version bump; other companies released some models too, but they barely made noise. It wasn't until around June or July that Kimi K2 started gaining traction, and it still couldn't beat R1 back then. One model dominated the conversation for half a year.

Overseas, Google's Gemini had been riding high since last fall. Nano Banana's text-to-image generation made quite an impact. I remember attending some top-tier academic conferences in September or October last year—everyone was talking about Gemini. Strong multimodal capabilities, good at coding, not weak in any category.

Then suddenly this year everything changed. After March, when talking about models, Google might as well have ceased to exist. Everyone's benchmark was Claude Opus, and then OpenAI caught up from behind. A few days ago I had the team run a blind test: OpenAI's Image Generation 2 and Nano Banana 2 each generated an image, and I deliberately didn't say which was which. The team all picked the OpenAI one. Turned out it was indeed OpenAI. Rumor has it Google's founders triggered another red alert over there, calling on all employees to charge ahead and catch up.

One quarter ago Google was in the "strongest, most comprehensive" position; one quarter later they're already in "we need to catch up" mode. That's how absurd it is. Advice you gave the team last time might need to be overturned and redone after three months.

The Winds Shift Too

Even funnier, even model styles are going through cycles of what goes around comes around.

My judgment a few months ago: GPT 5.4 was strong on complex tasks but didn't talk like a human; Claude 4.6 talked human but fell a bit short on some hard problems. So push Claude Code.

Using Opus 4.7 over the past two weeks, I found it's starting to not talk human. Unclear structure, sentences crammed with irrelevant fluff—much worse compared to 4.6's refreshing clarity. People in the community noticed this too, saying Opus 4.7 is starting to "GPT-ify."

Conversely, GPT-5.5 has started talking human again. Concise, readable, no rambling.

The two sides seem to have swapped identities.

This isn't just one model getting better or worse. It's the entire industry moving so fast that the half-life of your judgment is shortening. Three months ago I was quite certain we should push Claude Code; now I'm not so sure. It's not that anyone made the wrong choice—it's that "certainty" is something that's already hard to sustain in this industry.

The Best Bang for Your Buck

So I'm increasingly convinced of one thing: the best investment, bar none, is to immediately buy a coding plan account and start using it.

A few hundred dollars a month, or a few hundred RMB for domestic models. You get to use the world's most cutting-edge productivity at the earliest moment. And it upgrades every month and a half. Problems you can't solve today will likely naturally cease to be problems with the next generation in two months.

Why do I emphasize a coding plan instead of using the API directly?

Recent API experiments revealed that inside products like Claude Code and Codex, prompt caching hit rates are extremely high—the entire product chain is a coherent optimization pipeline. Caching, scheduling, and tool calls are all connected end-to-end. This is also why vendors can sustain such massive usage despite heavy losses. If you mess around with third-party API proxies and workarounds, this optimization chain breaks, making it cost-ineffective and prone to stability issues.

If you can, go straight for the official coding plan. You can open multiple accounts even—each one is only so much per month. Use each for their strengths.

Evolution Happens in Leaps

I've slowly figured the whole thing out: it's not linear.

It's not like 10 today, 11 tomorrow, 12 the day after. More like: 11, 12, 13, then stuck at 13 for a long time, then one day suddenly jumping to 100.

After o1 launched at the end of 2024, that winter I sat down with several frontier researchers to discuss. Everyone felt that the reasoning-plus-reinforcement-learning paradigm had a real shot at pushing models to the next scaling law. A few weeks later DeepSeek R1 was open-sourced. I still remember that shock. Top-tier results released directly as open source; R1 shot to the world-class tier the moment it came out, and reasoning became the new default paradigm.

Looking further back. GPT-4 came out in early 2023, o1 at the end of 2024—nearly two years in between was relatively flat. Honestly, in 2024 I had a sense of "slowdown." In 2023 everyone had their expectations fed too high by the leap from 3.5 to 4, and what followed couldn't keep up.

In retrospect that wasn't a slowdown; it was storing up energy. The next leap out was reasoning.

People who research AGI like to draw a staircase: conversation, tool use, agents, research, innovation. You think it's climbing linearly, but actually it wasn't working well for a long time, until one generation suddenly crosses a threshold. And once one crosses it, everyone can cross it.

Agents Aren't Designed—They're Pushed Out by Models

Another example of paradigm leap.

In early 2025, the entire community was discussing the same question: where exactly are AI applications? What applications do we build? No one could answer at the time. The mainstream was still IDE-embedded autocomplete (the Cursor route), vector database plus workflow orchestration (the LangChain route), and all sorts of drag-and-drop workflow tools.

Back then reasoning/thinking capabilities were seen as a burden by many teams. Task flows were already written, logic was fixed; the more the model thought, the easier it was to go off track. Many people recommended "turn off thinking and use a small model for higher efficiency."

By the Opus 4.5 generation (September/October 2025), context lengths grew, tool calling got stronger, and long-horizon autonomy emerged. Then Manus dropped a demo video, and everyone was stunned after watching it: models are already strong enough to push forward on long tasks by themselves. Some immediately said "there's no real technical difficulty here." And they weren't wrong—it really isn't hard. What's hard is realizing the models have reached the point where they can be used this way.

Later Claude Code took it even further. It's not an IDE embed, it's a CLI. A chat window in the terminal—you talk to it, it goes and does things itself. Many people at first couldn't accept "this can actually get work done?" After two days of use they realized: when models get strong enough, the IDE itself becomes a burden—clicking around inside it actually limits their performance.

Looking back now at that early-2025 question of "where are AI applications," it no longer exists. It wasn't solved; the question became meaningless. Models can do so much that the boundaries of the question dissolved.

Ordinary People Can Participate From Day One

One final thing I find particularly remarkable. This is something that ordinary people can participate in from the very first moment.

For technologies of this magnitude in the past—nuclear power, aerospace, semiconductors—it was impossible for ordinary people to touch the cutting edge within days of something new landing. AI is different. Before new models release, some insiders get early access a few weeks ahead, but the time gap is at most a month. Drag it out any longer and competitors catch up, diluting the lead.

So every time a new model drops, you and I can basically use it the same day. With an ordinary laptop, a few hundred bucks a month, you can judge how well it works, where it's strong, where it falls over. Sometimes you spot problems before the people who built it.

This sense of participation never existed before. The era itself is accelerating, and everyone caught in its rhythm gets pulled in.

A few hundred a month to stand at the frontier—I've never encountered such a good deal in my life.