Mars collides with Earth, a new model war!
Just 15 minutes after the release of Claude Opus 4.6, OpenAI also unveiled its latest and most powerful programming model.
GPT-5.3-Codex.
The most immediate feeling is that this new model finally has some aesthetic taste.
The official website showcased two demos: a racing game and a diving game. They're quite stylish.
It is said that GPT-5.3-Codex continuously iterated these games with almost no human intervention , accumulating millions of tokens .
In web development, in addition to better-looking UIs, there is also a stronger understanding of "intent".
Even if the prompt is unclear, it can automatically complete the logic and generate a fully functional website.
Judging from these demos, the design is indeed much better than before.
Its computer skills are also top-notch; it can now be used to help financial professionals create PowerPoint presentations directly.
It can also handle other workplace tasks, especially those requiring specialized knowledge; writing documents and creating spreadsheets are no problem for it.
In terms of hard power, the official highlights are as follows:
Smarter: SWE-Bench Pro 57%, TerminalBench 2.0 76%, OSWorld 64%.
More controllable: Supports real-time guidance during task execution, allowing for adjustments to the direction and updates at any time.
Faster: When completing the same task, less than half the number of tokens are required as in 5.2-Codex, resulting in a speed improvement of over 25% per token.
More Agent: Not only are they better at writing code, but they are also very strong in computer operations.
Looking directly at this comparison table will be more intuitive; almost every dimension shows a significant improvement over the previous generation.
Netizens exclaimed that it was too exciting, as OpenAI was just shot at by Anthropic with an ad yesterday, and today it retaliated.
Two heavyweight programming models in one day .
The comments section quickly split into the Anthropic faction and the OpenAI faction.
Let's take a look at how OpenAI performed in this AI coding war initiated by Ultraman.
GPT 5.3 Codex
What everyone cares about most is, of course, programming skills.
OpenAI states that GPT-5.3-Codex achieves state-of-the-art (SOTA) performance on SWE-Bench Pro .
This is a test designed specifically for real-world software engineering, covering four programming languages. It is more difficult, has more diverse tasks, and is closer to real-world production scenarios.
Meanwhile, GPT-5.3-Codex also shows a significant improvement in performance on Terminal-Bench 2.0.
More importantly, efficiency. While achieving these results, GPT-5.3-Codex used fewer tokens than any previous model .
Besides programming capabilities, another key focus of the new generation of Codex is computer use .
OSWorld is a computer usage benchmark for intelligent agents that requires models to perform various productivity tasks in a visualized desktop computer environment.
The results show that GPT-5.3-Codex is significantly stronger than the previous GPT model in terms of computer usage capabilities.
In summary, GPT-5.3-Codex is not a breakthrough in a single model capability, but rather a comprehensive development based on intelligent agents, with improvements in coding, front-end development, and computer operation .
What's even more interesting is that GPT-5.3-Codex directly participated in its own training process this time.
OpenAI stated that this is their first model to participate in "self-acceleration." The Codex team used an earlier version of the model during development to debug their own training process, manage deployments, and evaluate test results.
The official sources also provided some specific examples.
During the training phase , the research team used Codex to monitor and debug training tasks, helping to track changes in model behavior throughout the training process, conduct in-depth analysis of interactions, and propose improvement solutions.
In terms of data analytics , a data scientist collaborated with GPT-5.3-Codex to build a new data pipeline and visualize the results in a way that goes far beyond traditional dashboard tools.
The researchers then analyzed the results with Codex, and the model extracted key insights from thousands of data points in less than three minutes.
The engineering team then used Codex to optimize and adapt the testing and runtime framework for GPT-5.3-Codex.
When abnormal edge cases that affected user experience began to appear, the team members used Codex to locate the defect related to context rendering and further traced it back to the reason for the low cache hit rate.
Two More Things
The showdown with Anthropic was indeed quite exciting, but OpenAI actually has two other big moves worth noting.
1. Frontier: A platform that helps businesses create "AI colleagues"
This is a significant B2B business for OpenAI, with a clear goal: to truly integrate agents into the company's workflow.
Specific implementation methods include shared context, hands-on onboarding guidance, hands-on learning with feedback, and clear permissions and boundaries.
It is understood that well-known companies such as HP, Intuit, Oracle, State Farm, Thermo Fisher and Uber have already adopted Frontier.
2. AI4S: OpenAI and Ginkgo collaborated to reduce protein synthesis costs by 40% using GPT-5.
This is a laboratory-based company that does synthetic biology . They connected GPT-5 to an autonomous laboratory, allowing the model to propose experimental plans, execute experiments on a large scale, learn from the results, and decide what to try next, thus completing a closed loop.
2026 may be the year that AI4S evolves at an accelerated pace.
However, while OpenAI is busy clashing with Anthropic and netizens are dazzled by a series of new developments, there is another voice in the comments section.
Give me back my 40!
To this day, Ultraman has still not responded to the fact that Ultraman 40 has been completely removed from the market.
Perhaps, they were just too busy fighting Anthropic.
Reference link:
[1]https://openai.com/index/introducing-gpt-5-3-codex/
[2]https://openai.com/index/introducing-openai-frontier/
[3] https://x.com/i/trending/2019496485793198148
This article is from the WeChat public account "Quantum Bit" , author: focusing on cutting-edge technology, published with authorization from 36Kr.
