OpenAI launches a research preview of GPT-5.3-Codex-Spark, a smaller version of GPT-5.3-Codex that it claims generates code 15 times faster, for Pro users
This preview follows OpenAI's recent GPT-5.3-Codex release, which the company positioned as faster and better suited to longer-running work. Spark narrows that performance push toward interactive coding, where response delay can determine whether an assistant feels like a collaborator rather than a queued agent.
Related coverage also identifies Spark as OpenAI's first model running on Cerebras hardware, tying the product preview to a new inference-hardware deployment rather than a model-only update.
First-order effects
ChatGPT Pro users gain a research-preview coding model that OpenAI says substantially reduces code-generation and response latency, making short back-and-forth coding tasks more practical.
OpenAI segments its coding lineup: the larger GPT-5.3-Codex remains associated with longer-running tasks, while Spark is positioned for low-latency interaction.
Second-order effects
Coding-assistant rivals face greater pressure to compete on perceived responsiveness, not only code quality or autonomous task completion.
If Spark's speed depends on its reported hardware deployment, OpenAI gains a live test of whether specialized inference capacity can support interactive developer workloads at product scale.
Third-order effects
The split between a larger agent model and a smaller, faster conversational model points toward tiered AI coding stacks optimized for distinct workflow latency requirements rather than a single general-purpose assistant.
If this pattern holds, infrastructure choices will become more visible in AI product differentiation, as providers balance model capability, response time, and access tiers.
The trend: AI coding products are evolving from batch-style agents toward workflow-native assistants that pair task-specific models with latency-optimized inference.
GPT-5.3-Codex-Spark is launching today as a research preview for Pro. More than 1000 tokens per second! There are limitations at launch; we will rapidly improve.
BREAKING: @OpenAI just launched a new Codex model, Spark—it serves at 1,000 tokens per second. It's blow your hair back fast. It's their first model publicly released on Cerebras hardware, and you can see the difference. We've been testing internally @every for the last week or
GPT-5.3-Codex-Spark is the first milestone in our partnership with @cerebras. It provides a faster tier on the same production stack as our other models, complementing GPUs for workloads where low latency is critical. https://openai.com/...
Introducing GPT-5.3-Codex-Spark, our ultra-fast model purpose built for real-time coding. We're rolling it out as a research preview for ChatGPT Pro users in the Codex app, Codex CLI, and IDE extension. [video]
GPT-5.3-Codex-Spark size: ~700B@30B OpenAI's new GPT-5.3-Codex-Spark is the first model for which we can somewhat reliably estimate its size. Cerebras inference: 1000 tokens/s - GLM-4.7 is 355@32B, 92 layers 1400 tokens/s - Qwen3-235B is 235@22B, 94 layers 3000 tokens/s - [image]
ok this is very interesting, this is not the same perf than gpt5.3, and might not be the same arch as well? > Codex-Spark marks the first milestone in our partnership with Cerebras. Codex-Spark is optimized to feel near-instant when served on ultra-low latency hardware (from [ima…
Been using this model for a bit now, the combination of speed and intelligence is insane. It genuinely feels like a new paradigm shift. Excited to plug it into more specialized coding pipelines.
Hello GPT-5.3-Codex-Spark! ✨ Our first real-time coding model. It is... FAST. 1,000+ tokens per second. Once you experience latency this low, it's hard to go back. This is an exciting first milestone in our partnership with @Cerebras. [video]
This was so fun to work on. GPT-5.3-Codex-Spark built a snake game so fast that I was able to start setting high scores in about 9 seconds. If you have ChatGPT Pro you're going to want to check this out today! [video]
We've seen how much speed affects people's model preferences recently (e.g. the arena @swyx is running), so I think Codex Spark will be well-received. Also interesting that this initial release is a step towards combining long-horizon and real-time agents, including delegating to…
✨GPT-5.3-Codex-Spark✨ We're rolling out our first @cerebras model to Pro users today. It's fast! Rollout will be slow and very capacity constrained. Excited to roll out to more folks, and improve it with your feedback.
One of the top pieces of feedback we get about @OpenAI Codex: “make it faster”. We addressed it in a big way with our ultra fast Codex-Spark model (research preview for Pro). Available in the latest Codex app, CLI, and IDE extension. Here's Spark vs 5.3-Codex-Low side-by-side [vi…
GPT 5.3 Codex Spark! It's a smaller version of GPT 5.3 Codex running at over 1000 tokens per second on Cerebras hardware GPT 5.3 Codex was trained for GB200, on GB200. I wonder what this is? Maybe GPT 5.3 Codex that used Cerebras Reap?
I've been playing with GPT-5.3-Codex-Spark this week, and it's a really a✨ experience. Basically, using it for smaller tasks, context scanning, quick analysis, and smaller code edits. It feels so natural and instant; it's really hard to go back to other models.
Codex Spark was trained on GPUs for Cerebras hardware but OpenAI added support to their inference framework for Cerebras meaning they're reading to load future models onto it too GPUs are still foundational for inference and training, though [image]
Just one month after announcing our partnership with @OpenAI, we're launching our first model together: OpenAI Codex-Spark, powered by @cerebras. Codex-Spark is built for real-time software development. In coding, responsiveness is the product. It is not a nice to have. [image]