Inception AI Releases Mercury 3 Model, Claims 4x to 8x Speed Lead Over GPT-6 Luna and Haiku 5.5

Diffusion language model startup Inception AI Inc. has released its flagship Mercury 3 model, reporting that it outperformed OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 across eight of 12 internal benchmarks while running four to eight times faster. On Nvidia Corp. GPUs, Inception clocked Mercury 3 at 1,068 tokens per second, compared to approximately 240 tokens per second for Haiku 5.5 and about 130 for GPT-6 Luna. Rather than generating text sequentially one token at a time, Mercury 3 drafts an entire response upfront and refines every position in parallel over a few passes. While Inception did not disclose a parameter count, it stated the model is more than 40 times larger than Mercury 1 and features a 1 million-token context window, up from 260,000 in Mercury 2.5. In internal coding evaluations, Mercury 3 scored 51.1 on SWE-Bench Atlas compared to 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5, though it trailed Haiku 5.5 on the DeepSWE benchmark at 46.3 versus 51.6 and lost one long-context test. No outside organization has independently verified the benchmark results yet. Mercury 3 is priced at $0.20 per million input tokens and $0.70 per million output tokens—above the $0.10 input and $0.50 output rates of GPT-6 Luna and Haiku 5.5—though a launch promotion temporarily cuts both rates in half and cached inputs cost 90% less. Early adopters including smart-glasses search startup Octen AI and tutoring app RevisionDojo reported median generation times of 0.9 seconds to about one second via platforms including OpenRouter, Baseten, and Vercel. "The question the company has heard most since Mercury 1 is whether diffusion can scale. The limits come down to compute and data, not fundamental algorithmic barriers." — Stefano Ermon, Co-founder and Chief Executive of Inception AI