Changed
- Split the former single AI Speed table value into two clearer columns: AI Latency for benchmark wall-clock response time and TPS for rounded tokens per second. When a model's Health is not good, AI Latency now mirrors the exact Health text and TPS stays
—, so benchmark columns do not introduce a second, conflicting error language. - Made the AI benchmark prompt request one cohesive 80–100 word paragraph instead of a tiny one-sentence answer, and raised the benchmark cap to 140 output tokens. This gives latency and TPS measurements enough generated text to be more stable and meaningful.
Fixed
- Disabled reasoning/thinking for lightweight model ping payloads by sending
thinking: { type: "disabled" }on OpenAI-compatible chat-completion probes. This keeps latency checks focused on network/model availability instead of accidentally paying for hidden reasoning tokens. - Reused the same disabled-thinking ping payload for router health probes, while leaving Replicate prediction probes unchanged because they do not use the OpenAI chat-completions schema.