Anthropic and OpenAI have unveiled dramatically different approaches to speeding up large language model (LLM) inference, revealing a fundamental split in their technical strategies. Anthropic's 'fast mode' for Opus 4.6 offers 2.5x tokens per second, leveraging low-batch-size inference to reduce latency, though it still uses the full, expensive model. OpenAI's 'fast mode' for GPT-5.3-Codex-Spark achieves a staggering 1000 tokens per second – six times faster than Anthropic – but this relies on a significantly less capable distil model running on specialized Cerebras chips.

OpenAI's approach, while faster, sacrifices model fidelity, whereas Anthropic prioritizes using the actual, high-quality model at a premium cost. This divergence highlights contrasting priorities: Anthropic values model integrity, while OpenAI bets on hardware acceleration with a scaled-down model.