Inception AI announced the Mercury 3 model on Oct. 20, 2026. The company says its diffusion‑based approach can draft a whole response at once and then refine it in parallel, rather than writing one token after another. In its own tests Mercury 3 ran at 1,068 tokens per second on Nvidia GPUs, compared with about 240 for Claude Haiku 5.5 and 130 for GPT‑6 Luna.
What makes Mercury 3 faster?
The model first creates a rough draft of the answer, then improves every position in a few passes. This parallel refinement lets it generate longer outputs quickly. Inception did not publish the exact parameter count, only that Mercury 3 is more than 40 times larger than its predecessor Mercury 1. The context window grew to one million tokens, up from 260 000 in Mercury 2.5.
Why the speed matters to developers
Early adopters report noticeable improvements. Octen AI, which builds real‑time search for smart glasses, measured a median generation time of 0.9 seconds, saying a five‑second pause feels “broken.” RevisionDojo, a tutoring app, now delivers hints in about one second. Pricing is 20 cents per million input tokens and 70 cents per million output tokens, with a launch promotion halving those rates. By contrast, GPT‑6 Luna and Claude Haiku 5.5 list at 10 cents and 50 cents for typical prompts.
Why it matters
For developers building interactive products, the faster generation can make experiences feel more responsive, especially on devices like smart glasses where delays are noticeable. The lower cost during the promotion also lowers the barrier to try the model, though independent benchmarks are still pending.
