And the first OpenAI proprietary model not to run on NVIDIA GPUs is… GPT-5.3-Codex-Spark.
This leaner version of GPT-5.3-Codex indeed runs on the latest generation of Cerebras accelerator chips. Or, to be precise, on the era’s “megachips”: 4 trillion transistors across 460 cm², delivering 900,000 cores and 125 PFLOPS.
A low-latency complement to the GPU option
Cerebras – which also supply Mistral and Perplexity – had officialized in January its contract with OpenAI. The latter said it planned a phased integration into its inference stack, aiming to cover all of its workloads by the end of 2028. Not as a replacement, but as a complement to GPUs, to provide a very low-latency option (21 PB/s of memory bandwidth, 27 PB/s internal bandwidth, 214 PB/s inter-core). With GPT-5.3-Codex-Spark, this is expected to translate into, among other things, a throughput of over 1,000 tokens per second.
The model is currently in preview for ChatGPT Pro subscribers, on the Codex app, the CLI and the VS Code extension. Handling only the text modality with a 128k context window, it has its own usage limits, not deducted from the plan. OpenAI is also testing it on its API, with “a few partners.” Windsurf has furthermore integrated it into its leaderboard.
OpenAI to roll out WebSocket by default
The work aimed at reducing latency will benefit OpenAI’s other models as well. In particular, the introduction of a WebSocket pathway, enabled by default. Coupled with targeted optimizations on the API Responses and on session initialization, it has allowed, in the case of GPT-5.3-Codex-Spark, to reduce per-token overhead by 30%, for a TTFT (time to first token) halved.
On this basis, OpenAI intends to offer, in the longer term, a Codex “hybrid” that will exploit this type of model for simple interactions while knowing how to delegate the heavier tasks to sub-agents potentially running on GPUs.
We should recall that Sam Altman’s company also has an agreement with AMD, announced in October. It is set to translate into a 10% equity stake and the potential deployment of 6 GW of compute power – including 1 GW on GPU Instinct MI450 this year.
Further reading :
How OpenAI diversified its suppliers
Microsoft wants to fly on its own in AI
Why coding assistants don’t escape the productivity paradox
Claude builds its own C compiler: yes, but…
Agentic coding: Spotify’s learnings
