OpenAI's Codex Spark achieves 1,000 tokens per second on Cerebras

The GPT 5.3 Codex Spark model reached 1,000 tokens per second on Cerebras hardware, shifting the primary inference bottleneck from compute to network latency. To address this, the team implemented a persistent websocket mode to maintain stateful context and reduce overhead compared to standard HTTP server-sent events.

Featured · Dominik Kundel