OpenAI to Retire GPT-5.3-Codex-Spark Next Week
Tibo Sottiaux, head of Codex, announced that GPT-5.3-Codex-Spark will be retired next week due to a continuous decline in usage, and the company has significantly stronger models that need space for future products.
The model was launched on February 12 as a research preview, positioned as a smaller version of GPT-5.3-Codex, focusing on real-time coding. It was the first milestone of OpenAI's collaboration with Cerebras, aiming for near-instant feedback on wafer-level chips, with official claims of inference speeds exceeding 1,000 tokens per second, an order of magnitude faster than the flagship coding model.
Initially available only to ChatGPT Pro users, it covered Codex applications, command line, and VS Code extensions, with a context window of 128,000 tokens, and was pure text. The preview period had separate usage limits that did not count against the regular Codex quota; there could be queues during peak times. When cleaning the cluster in May, Tibo explicitly stated that Spark would not be retired at that time, only removing the entry points for GPT-5.2 and GPT-5.3-Codex in the ChatGPT login path.
The subsequent coding mainline has shifted towards larger context and higher capability tiers, including GPT-5.5 by default and GPT-6 Astra. While Spark's speed advantage remains, developers have concentrated long tasks and agent workflows on the updated models, leading to a decline in usage for this small model designed for low-latency interaction, with cluster management prioritized over retaining experimental models.
Tibo is also responsible for the core product platforms of Codex and ChatGPT, frequently announcing quota resets and model retirements. The Codex command line is written in Rust and open-sourced, with internal use of coding agents in code reviews, maintenance, and system restructuring; the product strategy is to merge ChatGPT and Codex into the same workflow rather than maintaining multiple generations of coding models in parallel.
From a market mechanism perspective, this is a reallocation of computing power rather than a disappearance of demand: buyers are shifting from immediate code modification scenarios for Pro users to new flagships capable of running long tasks; sellers are OpenAI reclaiming low-usage checkpoints from Cerebras' low-latency positions and GPU clusters. The event is driven by product line simplification, with funding and chip resources flowing towards higher-demand models like Astra. Beneficiaries include new model training and inference quotas, while workflows relying on Spark's feel for rapid iteration and Cerebras' non-NVIDIA real-time coding showcase are under pressure.
The collaboration with Cerebras made Spark the first production coding model fully running on non-NVIDIA hardware, emphasizing that GPUs and wafer chips can share the same workload. As flagship contexts expand to the million-token level, the 128k, pure text, Pro-limited preview tier loses its default entry position.
Source: Public Information
ABAB AI Insight
Tibo Sottiaux is not a customer service figure but the product head who transformed Codex from a research preview into the company's internal default engineering stack. His most notable external action is the hard reset of quotas, while internally he has open-sourced the Rust command line and embedded coding agents into reviews and restructuring. The birth of Spark corresponds to the January announcement of the Cerebras collaboration, delivering real-time performance of 1,000 tokens per second four weeks later, with the logic of using dedicated hardware to reduce interaction latency rather than stacking another set of homogeneous large models on GPUs.
The capital and resource mobilization is very direct: low-latency positions, independent quotas, and Pro entry points were used to validate whether "writing code can be as fast as typing." After the decline in usage, these resources were immediately redirected to GPT-6 Astra and higher default tiers. The motivation is cluster utilization—the coding product line has shifted from multiple checkpoints coexisting to a few flagship models plus APIs retaining old models. In May, the entry points for 5.2 and 5.3-Codex were cut first, while Spark was intentionally left, indicating that Spark's capabilities were not the issue, but rather that it could not compete with the default experience of new products even in terms of speed.
A similar path is seen in the rotation of multiple generations of models in GitHub Copilot, rapid iteration of Claude coding tiers by Anthropic, and Google merging Gemini coding endpoints into a unified studio. The industry is at a stage where coding agents are moving from a "model menu" to a "default workspace": users no longer reserve a checkpoint solely for speed but require context, tool calls, and repository-level tasks to be completed at once. Spark demonstrated that real-time tiers could be achieved but failed to become the default productivity layer.
The structural judgment belongs to technological substitution: when new models cover the interaction scenarios of old models within acceptable latency, dedicated acceleration tiers transition from products to transitional chip contracts. The mechanism is that both attention and quotas have limits—developers are only willing to remember one or two default models, and platforms are only willing to retain dedicated hardware for checkpoints with significant usage; once speed no longer constitutes an independent workflow, partner showcase positions will also be squeezed out of the product selector by the next round of flagship models.
ABAB News · Cognitive Laws
Speed is not a moat; default is.
The lifespan of preview products is determined by the latency of the next flagship.
Clusters do not nurture memories; they only nurture usage.