大模型
www.aibase.com ·
OpenAI 关停最快模型 GPT-5.3-Codex-Spark：发布到下线仅 7 个月
Codex 负责人 Tibo 发帖称 GPT-5.3-Codex-Spark 将于下周退役，该模型 2 月 12 日发布，生命周期约 7 个月。Spark 主打实时编程，1200 tokens/s、128k 上下文，是 OpenAI 首个脱离 NVIDIA 技术栈、基于 Cerebras WSE-3 晶圆级芯片的生产模型。但其 Terminal-Bench 2.0 仅 58.4%，远低于完整版 GPT-5.3-Codex 的 77.3%；SWE-Bench Pro 准确率停在 47%-51%，宣称的 15 倍加速在同精度下实际仅 1.37 倍，开发者还反馈其会编造 API 端点、JSON 格式不稳定。8 月 13 日 Cerebras 推出 Ultrafast 模式，将旗舰模型 GPT-5.6 Sol 分层分布到多个 CS-3 节点，在同等智能下达到 750 tokens/s，端到端加速 5.6 倍，被视为 Spark 被淘汰的直接原因。
AI推荐理由 内容提供了具体可核验的技术细节：GPT-5.3-Codex-Spark 于 2 月 12 日发布、约 7 个月后由 Codex 负责人 Tibo 宣布下周退役，1200 tokens/s、128k 上下文、Terminal-Bench 2.0 得分 58.4% 对比完整版 77.3%、SWE-Bench Pro 准确率 47%-51%、宣称 15 倍加速实际同精度下仅 1.37 倍，以及 8 月 13 日 Cerebras Ultrafast 让旗舰模型 GPT-5.6 Sol 直接跑在晶圆上实现 750 tokens/s 和 5.6 倍端到端加速。这些是判断专用高速小模型路线与 Cerebras 非 NVIDIA 算力落地价值的关键新信息。局限在于来源为 AIbase 转述，未给出 Tibo 原帖链接或官方公告，部分数字（如 750MW、200 亿美元订单）无法在片段内独立验证，且正文在‘Over the past thr’处截断，结尾信息不完整，因此质量分未给更高。
72
正文
OpenAI Shuts Down Its Fastest Model, GPT-5.3-Codex-Spark: It Lived for Only 7 Months from Release to Shutdown
AIbase基地
Published inAI News · 6 min read · Sep 14, 2026
10
OpenAI's fastest model ever is dead. Tibo, the head of Codex, posted that GPT-5.3-Codex-Spark will be retired next week. This model, capable of 1200 tokens per second, was OpenAI's first model to leave NVIDIA, and the first delivery of a 750 megawatt Cerebras order, lived only 7 months from release to shutdown. Tibo's reason was simple: usage has been steadily declining, and they already have significantly better models on hand, "it's time to make room for the future." He also added, "I can't believe we actually released a model with such a long name!!"
From 'future' to marginalization, defeated by 'dumb speed'
Spark was launched on February 12th with great fanfare: it was OpenAI's first model specifically designed for real-time programming, with a 128k context window, generating over 1000 tokens per second, and the official claim of an 80% reduction in round-trip cost and a 50% reduction in first token latency. It was also OpenAI's first production model running outside of the NVIDIA technology stack, with underlying computing power coming from Cerebras' wafer-scale chip WSE-3, marking the first delivery of that 750 megawatt, over $20 billion worth of computing power order.
But the hype faded faster than it rose. Spark's fatal flaw lay in the fact that a single wafer couldn't hold the flagship model; it was essentially a "distilled small model" compromised for extreme speed: in the Terminal-Bench 2.0 test, it scored only 58.4%, far below the full version of GPT-5.3-Codex's 77.3%; SWE-Bench Pro curves showed that although it could compress tasks to 1-2 minutes, its accuracy remained stuck at 47%-51%, while the full version could take 3 minutes to reach 51% and extend to 16 minutes to reach 57%. The claimed "15x speedup" was debunked - under the same accuracy, it was only 1.37 times faster. Developers criticized it for fabricating API endpoints out of nowhere and having unstable JSON formats. Turing College, a training institution, summarized it succinctly: "Speed without intelligence is just failing faster."
The real death sentence for Spark came on August 13th, when Cerebras released the Ultrafast mode: instead of a scaled-down version, the flagship model GPT-5.6 Sol ran directly on the wafer. By splitting the flagship model layer by layer and spreading it across multiple CS-3 nodes to form a pipeline, Ultrafast achieved 750 tokens per second under the premise of equivalent intelligence to the standard version - the same task took an average of 7.7 minutes on the standard version, but only 83 seconds on Ultrafast, achieving a 5.6 times end-to-end speedup. Cerebras CEO Andrew Feldman said, "Speed and intelligence are no longer mutually exclusive."
Speed goes from 'exclusive model' to 'paid tier'
Spark wasn't the only historical burden being cleaned up. Over the past three months, OpenAI's model library has undergone drastic changes: on June 2nd, GPT-5.2 and GPT-5.3-Codex were retired; on August 31st, GPT-5.4 and 5.4 Mini were retired, with users migrating to the 5.6 generation overall; on September 11th, it was GPT-5.3-Codex-Spark's turn. As the old generations left, Codex welcomed a new naming era with Sol, Terra, Luna, and Astra.
OpenAICodexGPT-5.3-Codex-SparkReal-timeProgramming
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -