Two different tricks for fast LLM inference

OpenAI and Anthropic have launched competing 'fast modes' for their top coding models, but with vastly different approaches. OpenAI's GPT-5.3-Codex-Spark reportedly achieves over 1000 tokens per second—a 15x speedup—using specialized 'monster' Cerebras chips. However, it's a less capable model. Anthropic's solution offers 2.5x faster speeds (up to 170 tokens/sec) using low-batch-size inference on its real Opus 4.6 model, but costs six times more for the latency reduction.