OpenAI Announced GPT-5.3 Spark, Which Writes Code in Real Time

OpenAI announced the GPT-5.3-Codex-Spark model, developed to increase speed in software development processes and minimize wait times. A product of the strategic partnership with Cerebras, this model promises real-time interaction by running on specialized hardware instead of standard GPU clusters.

Technical Features and Speed

GPT-5.3-Codex-Spark is positioned as one of the fastest code-focused models to date, with the capacity to generate more than 1,000 tokens per second (about 750 words). With a context window of 128,000 tokens, the model can analyze large code blocks in a single pass.

Delivering assertive results in benchmark tests such as SWE-Bench Pro, Spark is especially optimized for logic refactoring and targeted code edits.

Hardware and Infrastructure Improvements

Behind the model's high speed is the Cerebras Wafer Scale Engine 3 hardware. This specialized AI accelerator architecture enables the creation of a low-latency service layer. Performance gains are also supported by improvements made to the network infrastructure:

• Persistent WebSocket: Data transfer between client and server was accelerated by 80 percent.
• Latency Optimization: The dead time between the developer sending code and receiving a response was reduced to nearly zero.

Access and Use

For now, the model is offered primarily to ChatGPT Pro users under the “Research Preview” program. Available through VS Code extensions, CLI, and the Codex app, Spark has a speed limit policy different from standard usage limits due to its specialized hardware requirement. API access, meanwhile, is currently shared with a limited group of design partners.