Claude Haiku 5.5 Outscores GPT-6 Luna on Benchmarks, Yet Falls 6x Behind on Speed in Rocket Launch Test

Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna arrived within two weeks of each other in September–October 2026, both targeting cost-sensitive, high-volume workloads at an identical $0.10/$0.50 per-million-token rate. Yet a real-world performance test by Roundtable reveals a sharp divergence: Haiku leads on standard intelligence benchmarks, but Luna completed a rocket-launch-scene-building task in 1 minute 5 seconds for under one penny, while Haiku required 6 minutes 58 seconds and $0.05—a sixfold speed gap that underscores a fundamental trade-off in today's small-model race.