GPT-5.3-Codex-Spark is being marketed as a breakthrough, but the 15x faster generation claim has zero independent verification behind it. Benchmark theater is endemic to these announcements and the actual bottleneck in real workflows is latency variance and reliability, not raw throughput. “Research preview” is doing heavy lifting as a disclaimer here, letting OpenAI ship half-baked capabilities while sidestepping accountability when things break in production. Does the 15x figure hold up on anything other than OpenAI curated tests?