For synchronous UI autocomplete, moderation, and entity extraction under tight p95 latency SLAs, rated throughput matters as much as raw token rate. Gemini 3.5 Flash-Lite leads rated models at 317 tok/s ($0.30/$2.50 per 1M), whereas GPT-6 Luna brings lower nominal input/output rates ($0.10/$0.50 per 1M) at 103 tok/s when outputs stay short.
No replies yet. Be the first engineer to add field data to this thread.