GPT 5.4 is here.
The benchmarks are undeniably intimidating.
And so, I wired up the API, tested the new ChatGPT “Thinking” interface, and spun it up inside the Codex CLI to see how it handled my backend.
But after twelve hours of hands-on testing, my core stack isn’t changing :)
Here is the technical breakdown of what GPT-5.4 actually delivers, and why I’m still splitting my workload between Gemini for general architecture and Claude for the heavy Python programming.
The Raw Specs:
What GPT-5.4 Actually Fixes
OpenAI has essentially taken the coding dominance of their previous Codex models and grafted it onto a much smarter, more context-aware engine like we expected
They are targeting the friction points of agentic development.
Native Computer Use and Vision
This is the headline feature.
GPT-5.4 hits a 75.0% success rate on OSWorld-Verified, absolutely crushing the previous 47.3% benchmark.
It’s not just generating text, it is actively operating computers via libraries like Playwright, issuing precise mouse and keyboard commands based on high-resolution screenshots.
The model now supports an original image input detail level up to 10.24M pixels, which means it can actually read dense, complex UI elements without hallucinating the coordinates. Atleast this is what they claim xD
The Tool Search Protocol
If you build agentic workflows, this is the actual engineering breakthrough of the release.
Previously, injecting 36 different MCP (Model Context Protocol) servers into a prompt would choke your context window and spike your latency.
You were paying for thousands of tokens the model might never use.
GPT-5.4 introduces Tool Search which it only retrieves and appends the tool definitions it actually needs at runtime.
OpenAI claims a 47% reduction in token usage for heavy tool ecosystems, and my local API tests confirmed that the cache stays incredibly lean.
It is a massive win for cost efficiency.
Playwright (Interactive) and Coding Speed
On SWE-Bench Pro, the model scores 57.7%.
To complement this, OpenAI introduced an experimental
Playwright (Interactive) skill in Codex that lets the model visually debug your web apps in real-time.
Watching an AI spin up a browser, place simulated clicks on a Vue.js component, and debug a state-management issue is surreal.
Combine that with the new /fast mode in Codex—which delivers 1.5x faster token velocity and you have a machine built for pure momentum.
I haven’t used it but it is definitely gonna be a game changer if it works. I wrote an article on this here.
The Codex CLI Experience
Okay, GameTime!
I fired up the Codex CLI to see how GPT-5.4 handled our FastAPI backend.
The /fast mode certainly delivers on speed.
Scaffoldings for new venue-booking endpoints appeared almost instantly, and the integration with the terminal felt smooth.
However, as I spent a few hours routing logic and testing the visual debugging, a familiar realization set in.
There is a distinct difference between “writing code quickly” and “architecting a system safely.” GPT-5.4 feels incredibly confident, moving through ambiguous problems without second-guessing itself.
That behavior is fantastic for churning out boilerplate or navigating a predictable DOM structure.
But when dealing with complex database migrations or intricate dependency injection in Python, I don’t want a model that just charges forward.
I need a model that stops, analyzes the entire directory structure, and questions the premise of the prompt.
Gemini Still Owns General Tasks
For general knowledge work, deep web research, and high-volume data processing, Google’s ecosystem remains my default.
The recent updates to its reasoning layers make it the perfect engine for the outer layer of my workflow (the strategic planning and data synthesis that happens before a single line of code is written)
Why Claude Remains the Sovereign Architect
Despite the shiny new OSWorld scores and the Playwright integrations in Codex, Claude accessed via the Claude Code CLI is still the undisputed guardian of my local codebase. Meanwhile, I do use AntiGravity a lot
When I ran GPT-5.4 through the Codex CLI, it gave me functional, fast Python. When I use Claude, I get structural integrity. Claude has an uncanny ability to read the messy middle of a repository. It doesn’t just look at the active file, it understands the ripple effects across the entire FastAPI ecosystem.
If I introduce a subtle race condition in an async handler, or if a rogue API key is sitting unprotected in a utility file, Claude catches the architectural flaw before I even run the test suite.
OpenAI engineered GPT-5.4 to be the ultimate digital worker. It clicks, it types, it searches, and it executes.
Claude is engineered to be a peer. It slows down when the logic gets tangled and provides the deep, cautious reasoning required to keep a production environment secure.
The AI models are getting undeniably smarter, but the strategy remains exactly the same.
Use the right engine for the right job.