OpenAI Retires Codex-Spark, Its Fastest Model, After Just 7 Months
新智元 · wechat · 2026-09-12
OpenAI Codex lead Tibo announced GPT-5.3-Codex-Spark will be retired next week, only seven months after launch. The model generated over 1,200 tokens per second and was OpenAI's first production model running outside the NVIDIA stack, powered by Cerebras wafer-scale WSE-3 chips — the first delivery under a 750 MW, $20B+ compute deal.
Usage kept declining, and the distilled speed-first model scored just 58.4% on Terminal-Bench 2.0 versus 77.3% for full GPT-5.3-Codex; the claimed 15x speedup was debunked (only 1.37x at equal accuracy). Developers valued its separate quota more than the model itself, using it as a backup tank when Astra limits ran out.
The death blow came in August when Cerebras' Ultrafast mode ran the full flagship GPT-5.6 Sol at 750 tokens/sec with no intelligence loss — 5.6x end-to-end faster. Speed is shifting from dedicated models to paid tiers on flagship models, as OpenAI retires old models and moves to its Sol/Terra/Luna/Astra naming era.
Related event: OpenAI retires speed-king Codex-Spark after just 7 months(2 posts)→
More from Infra
- High Refractive Index Glass Hits 2.1 With 97% Transmittance, Raising Eyebrows for AR Optics — jwt0625 · 2026-10-24
- EPFL demos ~100GHz-spacing TFLT WDM modulator chip; expert doubts CPO fit — jwt0625 · 2026-10-24
- Reddit deep-dive: next-gen AI via memristors, spiking nets, and 1% sparse routing — Robert__Sinclair · 2026-10-24
- ai& cuts MiniMax H3 video generation from 229s to 19s across NVIDIA, AMD, Tenstorrent — DavidBennett__ · 2026-10-24
- Local AI benchmarks miss thermal throttling — a stand can boost NVIDIA Spark perf — draginol · 2026-10-24
- Cloudy: autonomous cloud research agent hand-writes Triton kernels that beat tuned vLLM by up to 15% — saurabhtwq · 2026-10-24