GPT-5.3-Codex represents a brand new technology of the Codex mannequin constructed to deal with actual, end-to-end work. As a substitute of focusing solely on writing code, it combines sturdy coding capability with planning, reasoning, and execution. The mannequin runs sooner than earlier variations and handles lengthy, multi-step duties involving instruments and choices extra successfully.
Quite than producing remoted solutions, GPT-5.3-Codex behaves extra like a working agent. It might keep on process for lengthy intervals, alter its method mid-way, and reply to suggestions with out shedding context.
Codex 5.3 Benchmarks
OpenAI’s GPT-5.3 Codex units new efficiency requirements on real-world coding and agentic benchmarks, outperforming prior fashions on exams like SWE-Bench Professional and Terminal-Bench 2.0 with stronger accuracy. It additionally reveals substantial positive factors on OSWorld and GDPval evaluations, which measure computer-use {and professional} information work, whereas operating about 25% sooner than GPT-5.2 Codex. This marks a big step towards AI that may deal with longer, multi-step improvement duties and broader software program workflows.
Key Options
Right here’s what makes OpenAI Codex attention-grabbing:
Constructed With Codex, For Codex
One of the crucial attention-grabbing elements of GPT-5.3-Codex is that the staff used early variations of the mannequin throughout its personal improvement. Engineers relied on it to debug coaching runs, examine failures, and analyze analysis outcomes. This helped pace up iteration and uncovered points earlier within the course of.
This self-use is a powerful sign of maturity. The OpenAI staff not solely examined the mannequin on benchmarks but in addition trusted it in actual inner workflows.
From the benchmark picture, we are able to see that GPT-5.3-Codex maintains increased accuracy as output tokens improve. It performs higher on longer and extra advanced duties. This reveals stronger consistency in comparison with earlier fashions.
Anthropic additionally launched their new coding mannequin lately. Discover all about it on our detailed weblog on Claude Opus 4.6.
Past Writing Code
GPT-5.3-Codex is designed to deal with extra than simply code technology. It might assist with debugging, refactoring, deployment duties, documentation, information evaluation, and even non-coding work like writing specs or making ready stories.
It operates greatest when given targets moderately than detailed directions. The mannequin can resolve what to do subsequent, run instructions, examine outputs, and preserve going till the duty is full.
Designed for Protected, Sensible Use
To assist hands-on work, GPT-5.3-Codex runs inside managed environments. By default, it really works in sandboxes that restrict file entry and community utilization, decreasing the danger of unintentional injury. The mannequin additionally pauses and asks for clarification earlier than performing probably damaging actions.
These decisions make it simpler to experiment, particularly when engaged on actual initiatives or unfamiliar programs.
Working Collectively With the Mannequin
Interplay with GPT-5.3-Codex is steady moderately than one-off. As it really works, it shares progress, explains choices, and reacts to suggestions. You’ll be able to interrupt, redirect, or refine the duty at any level.
This makes it really feel much less like a command-based device and extra like a collaborator you supervise.
Easy methods to Entry Codex 5.3?
Now that the high-level image is evident, it’s time to maneuver from description to motion.
Within the subsequent part, we’ll attempt Codex hands-on. We’ll begin by downloading and setting it up, then stroll by a easy workflow step-by-step. This can present how GPT-5.3-Codex behaves in apply and learn how to work with it successfully on actual duties.
Let’s see the steps:
1. Drag the Codex icon into your Software folder
2. Open Codex
3. Register with ChatGPT
4. After signing in, choose a folder or git repository in your laptop the place Codex will work
5. Kick off your first process
5. Choose the mannequin from right here and Reasoning as per your selection.
Process 1: Textual content to 3D Scene Generator
The primary process I labored on with Codex was constructing a easy text-to-3D scene generator. The objective was deliberately minimal. I needed to check how nicely Codex may take a loosely outlined concept and switch it right into a working visible venture with out overengineering.
The Preliminary Immediate
The very first immediate I gave Codex was easy:
Construct a easy text-to-3D scene generator.
The necessities had been clear however restricted. It needed to be a single HTML file, use Three.js by a CDN, and run instantly within the browser with no construct instruments. The scene wanted a textual content enter the place a person may describe one thing like “3 bushes and a home”, and the output needs to be a fundamental 3D scene utilizing easy shapes, lighting, and gradual rotation. I additionally requested it to begin with a minimal working model.
This immediate was meant to check fundamentals, not polish.
First Working Model
Codex created a clear index.html from scratch. It arrange a Three.js scene with a digicam, lights, floor aircraft, and a easy animation loop. A textual content enter and submit button had been added. Primary key phrases like tree, home, cloud, and solar had been parsed and mapped to easy shapes. The main target was correctness. The scene loaded, objects appeared, and all the pieces rotated easily. The outcome was already usable.
Iterations
I iterated step-by-step. I improved parsing so phrases like “3 bushes” labored accurately, with a default of 1 object. Subsequent, I mounted object spacing to stop overlap and added scene cleanup so every submission rebuilt the scene as a substitute of stacking objects. In one other go, I centered on readability by simplifying feedback and clarifying the construction for rookies. Every change was small and fast to implement.
Outcome
By the third model, a number of objects rendered accurately, however it took extra time than anticipated and the outcome was nonetheless not very sturdy. The scene did clear and rebuild on each submit, however the conduct was inconsistent. Within the video, you can too see that after I entered “cone,” nothing modified within the scene. The ultimate output ran within the browser, however it clearly confirmed that Codex may do extra and that the answer was removed from its full potential.
Process 2: Area Flight Sandbox
This process centered on constructing a real-time house flight sandbox with a powerful emphasis on construction and efficiency. The objective was to create a easy and plausible expertise the place the system may scale with out breaking.
Core Gameplay
The participant flies a ship in open house with inertial motion. Mouse enter controls pitch and yaw, whereas the keyboard handles thrust, strafe, roll, and reverse. A big asteroid subject surrounds the participant and constantly streams because the ship strikes. The participant can fireplace lasers to destroy asteroids, which cut up into smaller items when hit.
Efficiency and Construction
Efficiency was handled as a tough constraint. Asteroids had been rendered utilizing InstancedMesh and recycled to take care of a steady occasion rely. Collision checks relied on a spatial grid to remain environment friendly. Physics ran on a set timestep, whereas rendering remained easy and decoupled. No exterior physics engines or frameworks had been used.
System Design
The venture adopted a clear modular design. Every main system lived in its personal file, with most important.js dealing with the scene and loop, ship.js managing flight physics, asteroids.js dealing with instancing and streaming, weapons.js managing lasers and collisions, and controls.js dealing with enter. This construction remained unchanged all through improvement.
Audio Suggestions
Audio was added to enhance readability and impression. Laser pictures set off a pointy firing sound, and asteroid hits play a heavier explosion-like thud. All audio makes use of Three.js Audio and is hooked up to the digicam to remain in keeping with the participant’s perspective.
Outcome
The ultimate sandbox is absolutely playable and steady, however it took for much longer to construct than anticipated. The ship feels weighty and responsive, asteroids stream endlessly with out efficiency drops, and lasers really feel highly effective and visual. Nevertheless, the event time was noticeably excessive, probably because of the reasoning mannequin I selected. After seeing the outcome, I used to be not very proud of it, as different fashions do significantly better, or this might have been made significantly better total.
Conclusion
GPT-5.3-Codex reveals clear strengths in lengthy, advanced duties and benchmark efficiency. It behaves extra like an agent than a easy code generator. It plans, executes, and adapts over time. Benchmarks recommend sturdy consistency at scale. Nevertheless, hands-on work revealed gaps. Some duties took longer than anticipated. Outcomes weren’t all the time as sturdy as they may have been. In apply, iteration pace and output high quality diverse. Whereas the mannequin is highly effective and mature, the workflow didn’t all the time really feel optimum. With higher decisions or tuning, the identical duties may seemingly be executed sooner and higher.
Hello, I’m Janvi, a passionate information science fanatic at present working at Analytics Vidhya. My journey into the world of information started with a deep curiosity about how we are able to extract significant insights from advanced datasets.
Login to proceed studying and revel in expert-curated content material.
