GPT 5.4 Made History in 13 Seconds

A complete package (utter garbage for some)

This time it’s GPT 5.4. And honestly it feels less like a breakthrough and more like OpenAI finally figured out what they should have been building all along.

The marketing around 5.4 was explicit: combine the coding capabilities of Codex 5.3 with the knowledge work and web search skills of GPT 5.2. Make the all-rounder. The do-everything model that doesn’t suck at anything.

According to third-party benchmarks, they actually achieved it.

It ranks as the best coding model. The best agentic model. It ties with Gemini for best intelligence. It’s not the flashiest release. It’s not the most interesting architecturally. But it might be the most competent model OpenAI has ever shipped.

And that’s a problem.

What 5.4 Actually Does.

The headline features are straightforward: better at knowledge work and web search. Native computer use capabilities. Tool search. Can be steered mid-response. New fast mode. One million token context window.

On paper it’s just taking features from previous models and combining them.

But the computer use capability is genuinely interesting. OpenAI designed this as their first general purpose model with built-in computer use. It can write code using libraries like Playwright. It can issue mouse and keyboard commands in response to screenshots. It can see the screen and actually manipulate it.

It’s not perfect. But it’s genuinely hands-off development.

You prompt it. It builds. You verify. It improves.

That’s a different category of tool.

The Boring Upgrades That Actually Matter.

But here’s what’s actually interesting about 5.4: it’s mostly boring.

Tool search solves a real problem that nobody talks about. If you have 36 MCP servers, you end up loading all their tool definitions into the system prompt upfront. That’s wasteful. That’s context bloat. That’s fewer tokens for your actual task.

With 5.4, the model gets a lightweight list of tools. When it needs a tool, it searches for it and loads just that tool definition right when it’s needed. OpenAI claims this reduces token usage by up to 47 percent while maintaining accuracy.

That’s not the kind of thing that gets shared on Twitter. But it’s pragmatic. It’s the kind of optimization that makes systems actually work at scale.

The model also got better at using tools. Better at deciding when to use them. Better at completing complex tasks that require multiple tool uses in sequence.

Again. Not flashy. But important.

Fast mode is another boring upgrade that’s actually clever. It’s the exact same model. Same intelligence. Same experience. But delivered up to 1.5 times faster because you’re paying for priority compute. It’s a tier not a different model.

After spending two hours with 5.4, it told me I could have saved an hour if I’d switched to fast mode.

That’s both funny and telling lmaooooo.

The Real Strength.

The actual strength of 5.4 is that it doesn’t have a weakness.

Previous models had trade-offs. Codex was fast and good at coding but sometimes struggled with broader knowledge work. Opus was careful and thoughhtful but slow. GPT 5.2 was good at knowledge work but less impressive at sustained agentic tasks.

5.4 is just competent at everything.

That’s boring but who the f gives a flying f

Smarter. Runs for longer. Uses tools better. Completes harder tasks. That’s the feature list.

The Cons

But 5.4 has some real issues that nobody’s talking about because they’re focused on benchmarks.

The first is speed. And I mean that in two ways. The model thinks for a long time before outputting. Sometimes this is good because it’s reasoning through things carefully. Sometimes it feels like it’s just slow.

Artificial Analysis data shows 5.4 takes the longest to return a token by a significant margin compared to other models. And the same applies for returning the first 500 tokens.

I don’t know if it’s a model issue or a provider issue but it’s noticeable tho

The second is price. The base model is 2.50 per million input tokens and 15 per million output tokens. Fine. But the Pro version is 30 per million input and 180 per million output. And if you use the new one million token context window, anything beyond 272,000 tokens gets billed at double the normal rate.

That’s EXPENSIVE.

If you’re using the extended context you’re paying a premium. If you’re using the Pro version you’re bleeding money.

This model is built for people with serious API budgets, so yeah, not you and me (unless you’re some rich guy reading, in that case, good for you dude)

The third is UI design and this one is subjective but important. When I asked 5.4 to design a cafe website compared to Opus 4.6, I preferred Opus. Neither blew me away. But 5.4 leans into this frosted glass card aesthetic with gradients that all OpenAI models seem to share.

It’s formulaic.

On Design Arena, this model doesn’t rank highly. OpenAI just isn’t strong at creative design right now. And for a model that claims to be the all-rounder, that’s a gap.

Why This Is Actually the Most Important Release.

Here’s what’s actually happening: OpenAI is converging toward competence.

They’re not building bleeding-edge specialized models anymore. They’re building mid-range models that do everything decently. Codex was specialized for coding. Opus is specialized for reasoning. But 5.4 is just competent at everything.

That’s a different market position.

And it’s a threat to the entire model ecosystem.

Because most people don’t need the best coding model or the best reasoning model. They need a model that’s good enough at everything and doesn’t require them to think about which tool to use.

5.4 is that model.

You don’t have to choose between speed and depth. Between coding and reasoning. Between web search and knowledge work. It just does everything.

For most use cases, that’s the right choice.

Which Model to Actually Use.

If you’re already deep in the Codex workflow, upgrade to 5.4. It’s faster. It has better tools. It’s computer use is genuinely useful. The speed issue and the price issue are real but they’re worth it if you’re shipping code.

If you care about creative design work, stick with Claude. Opus is better at this. GPT models all look the same.

If you’re doing knowledge work and web search, 5.4 is probably better than anything else. The one million token context is useful. The tool search is useful. The speed tradeoff might be worth it.

If you’re price sensitive, don’t use 5.4. It’s expensive. Use something cheaper or wait for prices to normalize.

The Real Question.

The actual interesting question isn’t whether 5.4 is better. It obviously is. The question is whether generic competence beats specialized excellence.

5.4 is better at coding than GPT 5.2 but worse than Codex. It’s better at reasoning than Codex but worse than Opus. It’s better at design than Opus but still not great.

For most people, that trade-off is fine.

For people with specific needs, you might still want the specialized model.

But the trend is clear.

OpenAI is betting that most people want one model that’s good at everything instead of different models for different things.

And they might be right.

Because switching models is friction. Managing different contexts is friction. Figuring out which tool to use for which task is friction.

A model that just handles everything eliminates friction.

That’s worth something.

So yeah. Daily driving 5.4.

The boring upgrade that’s actually important.

Want more stuff like this? Even more detailed? I go into deep dives every single day on my 📌Substack📌

At the price of a coffee I can give you enough knowledge to replace a bachelor’s degree in computer science or maybe even a masters, who knows.

💜 If you like my work and would like to support to me financially. Even a small donation would help a lot!

My PayPal — Please Support