The emphasis on the Codex of OpenAI increases in 2026. The artificial intelligence (AI) powerhouse based in San Francisco published the GPT-5.3-Codex earlier this month, the first model where the company prioritized a coding-oriented model, rather than the general-purpose variant. The growth of the Codecs was reportedly termed as insane by the OpenAI CEO Sam Altman, last week because it increased by 50 percent within a week. Today, still on this trend, the AI giant has published its first AI model capable of writing code in real-time, named GPT-5.3-Codex-Spark.

OpenAI Introduces GPT-5.3-Codex-Spark
On a post, OpenAI commented on its latest coding model and described it. GPT-5.3-Codex-Spark is already in research preview on the Codex app, ChatGPT Pro subscriber extension, command line interface (CLI), and an integrated development environment (IDE) extension. Users of the model will be limited at the moment since the full version of the model has not been introduced.

In addition to its unusual name, which is Codex-Spark, the codex has the most remarkable feature of real-time coding. According to the company, the model is tailored to the low-latency workload and can write and edit the code nearly immediately. It is said that the model can process 1,000 tokens per second.

It is a text-based model that is able to write code in real-time, add and remove targeted changes, change logic, and enhance the interface. It has a context window of 128,000 tokens, and can be used to accommodate tasks of a routine to moderately complex nature. Nevertheless, when tackling more challenging problems, the GPT-5.3-Codex model should be used.

In addition to optimizations at the code level, one of the reasons why Codex-Spark can be faster is that the AI model has a low-latency hardware implementation. In it, OpenAI also partnered with Cerebras last month, and as such, the new model is now operating on its Wafer Scale Engine 3 AI accelerator, which offers high-speed inference.

About performance, the AI giant gave internal benchmark results, stating that on the SWE-Bench Pro and Terminal-Bench 2.0, which test agentic software engineering performance by measuring several aspects, including, but not limited to, the use of internal memory, Codex-Spark outperforms GPT-5.1-Codex-mini and slightly underperforms GPT-5.3-Codex. This is a big step towards advancement with increased output generation.

OpenAI indicated that the AI ​ model will also be opened to select design partners through an application programming interface (API) to ensure that the firm can be aware of how developers are interested in incorporating it in their products. Notably, Codex-Spark will possess its rate limits and will not incur usage on standard rate limits. The company indicated that this would be extended within the next few weeks.