Course
Working with the OpenAI API
3 hr
103.1K
In sports, you would call this a counterattack. Just 30 minutes after Anthropic published their new Claude Opus 4.6 model, OpenAI released a major update as well.
Their new GPT-5.3-Codex model replaces both GPT-5.2 and GPT-5.2-Codex. Its main focus is on combining the strengths of these two legacy models to provide a more general agentic experience. In combination with the Codex app for macOS, introduced only a few days earlier, it also enables interactive, real-time collaboration without the risk of losing context.
In this article, we will cover all the new features, take a look at the benchmarks, and see how GPT-5.3-Codex works in a couple of hands-on examples. We will also try to examine how well the model actually performs and how it compares to Anthropic’s Claude Opus 4.6.
If you are interested in learning more about OpenAI’s latest features, I recommend reading our guides on ChatGPT Images and ChatGPT Health.
What is GPT 5.3 Codex?
GPT-5.3-Codex is OpenAI’s newest large language model (LLM), following up on GPT-5.2 and GPT-5.2-Codex, which were both released in December 2025.
In contrast to these two legacy models, the new release takes a new approach. While there was a clear distinction between coding agent and reasoning LLM in the GPT-5.2 models, GPT-5.3-Codex merges them and is introduced as a general-purpose agent excelling at both.
The GPT-5.3-Codex model is supposed to not just write functions, but also understand the work around the code. Think of updating Jira tickets, writing documentation, or managing deployment pipelines.
Performance-wise, the new model almost doubles its score in the OSWorld-Verified benchmark and sets new high scores for both SWE-Bench Pro and Terminal-Bench. Additionally, OpenAI focused on efficiency and claims that the new model will be 25% faster due to improvements in infrastructure and the inference stack.
One notable thing is that OpenAI apparently used GPT-5.3-Codex to actively debug and manage its own creation. While other frontier models like Gemini 3 generated their own training data, Codex went a step further by acting as a site reliability engineer: monitoring its own training runs, diagnosing infrastructure errors, and writing scripts to dynamically scale GPU clusters during launch.
Key Features of GPT 5.3 Codex
The release of GPT-5.3-Codex focused on enabling general agentic workflows. Let’s take a look at some key features.
The general work agent
In contrast to its Codex predecessor, GPT-5.3-Codex is designed to be a general work agent. The aim is to transcend the IDE, with the model effectively handling “knowledge work” alongside “coding work.”
The new model is built to support all work across the software lifecycle:
- Engineering and operations: Handling the technical "heavy lifting" like debugging, testing, deploying, and ongoing monitoring of systems.
- Product and planning: Supporting the strategic side of development by writing product requirements documentation and assisting with user research.
- Analysis and communication: Managing the "soft skills" of software delivery, including editing copy and tracking project metrics.
This versatility enables GPT-5.3-Codex to execute end-to-end workflows. The model could, for instance, write an SQL query, fetch the data, and then generate a PDF report or slide deck based on it via tool calls.
Interactive real-time collaborator
The interactive collaborator feature is the biggest perk of the Codex app and has the potential to make the biggest difference in everyday work. It keeps you in the loop throughout the process and lets you intervene in real time.
Essentially, GPT‑5.3-Codex constantly lets you know what it is doing and offers you the chance to steer it in the right direction long before you receive the final output. Instead of waiting, you can ask questions, give feedback, or add context to your initial prompt. The model then responds to your feedback and adapts mid-stream.
Currently, the Codex app is only available for macOS. You can turn on steering in the app settings under General > Follow-up behavior.
Cybersecurity focus
OpenAI also shifted its focus to cybersecurity, particularly to vulnerability detection. GPT-5.3-Codex is the first model classified as "high capability" under OpenAI’s Preparedness Framework, meaning it is specifically trained to identify and fix software vulnerabilities.
To balance this power with safety, OpenAI has deployed a defensive stack designed to prevent misuse, such as automating cyberattacks. It includes safety training, real-time monitoring, and Trusted Access for Cyber, a pilot program that gates advanced capabilities to verified researchers.
Furthermore, OpenAI is investing heavily in the ecosystem, launching the Aardvark security agent (currently in beta) and committing $10M in API credits to support open-source maintainers with free code scanning tools.
GPT 5.3 Codex Benchmarks
While we are still waiting for verified results in many of the state-of-the-art benchmarks, the announcement featured scores in several areas:
- Agentic workflows: OSWorld-Verified
- General coding: SWE-Bench Pro
- Agentic coding: Terminal-Bench 2.0
- Reasoning: GDPval
Agentic workflows
OSWorld-Verified is the gold-standard benchmark for testing an AI's ability to operate a computer like a human. It goes beyond simple text processing by placing the AI in a real virtual machine and asking it to complete open-ended tasks using a mouse, keyboard, and GUI apps (e.g., "Open LibreOffice, create a spreadsheet with this data, and save it as a PDF").
GPT-5.3-Codex achieves 64.7% in the OSWorld-Verified benchmark. That’s a staggering increase of 26.5 percentage points compared to its predecessor, GPT-5.2-Codex. This strong result reflects OpenAI’s focus on creating a more general, agentic experience for GPT-5.3-Codex, optimized for good performance across tasks and domains.
Coding
Software development was the initial focus of the Codex models. On the SWE-bench Pro (Public), GPT-5.3-Codex reaches 56.8%, only a minor increase from 56.4% with GPT-5.2-Codex. The incremental improvement here is likely the trade-off made in optimizing for agentic skills.
On the agentic coding side, we can see a quite significant jump: GPT-5.3-Codex scores 75.1% on Terminal-Bench 2.0, a substantial increase from the 64% with GPT-5.2-Codex. Even more interesting, it topped the result of Claude Opus 4.6, which had claimed to top the benchmark just half an hour earlier, by over 5 percentage points!
Reasoning
For the model’s reasoning skills, there’s not really anything exciting to report. GPT-5.3-Codex reaches exactly the same result as GPT-5.2 on GDPval (70.9%). It’s fair to interpret this in a way that the (good) reasoning skills of GPT-5.2 were incorporated into the Codex model, without focusing on substantial improvement in this area.
