Inception launches diffusion language model Mercury 3, claims 4-8x speed of GPT-6 Luna

Mercury 3 drafts a whole answer and then refines it over multiple rounds rather than generating token by token, hitting 1,068 tokens per second on Nvidia GPUs, versus roughly 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna. It handles up to 1M tokens at once; parameter count is undisclosed, though it is said to be over 40x Mercury 1 with no speed loss. It won 8 of 12 Inception-run benchmarks: 76.84 on IFBench, 15 points ahead of GPT-6 Luna, and 51.08 on the SWE-Bench Atlas autonomous coding test versus 33.33 and 32.26. But Claude Haiku 5.5 won the long-coding DeepSWE and long-context retrieval tests, and GPT-6 Luna led two Artificial Analysis tests. Initial targets are agents, search, customer service and tutoring, where every step waits on the model. Smart-glasses search partner Octen AI says the language model was the bottleneck and now sees a 0.9s median response and 2.2s end to end; education firm General Learning says hints and corrections arrive in about 1 second; TinyFish AI says it solved 25 of 50 real-website navigation tasks. Pricing is $0.20 per million input tokens and $0.70 per million output tokens, at 50% off during launch with no stated end date; after the discount, output pricing on short inputs will exceed rivals' $0.50, though rivals charge far more on long prompts. The model is live on Inception's own platform plus OpenRouter, Baseten and Vercel, with an OpenAI-compatible API requiring minimal changes to switch.