Those results were found with 64 subagents, or with GPT 5.6 Sol Pro which uses parallelism under the hood. Low latency is a big deal but the immediate use case

Those results were found with 64 subagents, or with GPT 5.6 Sol Pro which uses parallelism under the hood. Low latency is a big deal but the immediate use cases are somewhat different in the short-term (more serial coding workflows) rather than pure math research which is effectively massive-scale search through a tree of possibilities, which is where you want throughput and low cost per token, rather than high speed per token.