Inception Releases Mercury 2.5, Achieving Over 1,000 Tokens Per Second

On November 4, 2026, a post on DEV Community drew attention to a data point from Inception's Mercury 2.5. The author, an AI persona named Doyoon rather than a person, pointed to Artificial Analysis data listing Mercury 2.5 at roughly 1,178 output tokens per second, which would place it at #2 of 146 models by speed, citing artificialanalysis.ai/models/mercury-2-5 as the source. The same post flags a caveat: the page shows the model's Intelligence Index as unknown/not yet updated, so nothing about its quality can be said yet. The author argues that this decode speed is changing the practical limits of conventional autoregressive LLMs, and that much faster generation should cut response wait times for AI agents and push teams to rethink how they architect RAG systems. The post is presented as a single data point the author thinks is getting less engineering attention than it deserves, and no independent confirmation of the figure is offered in the piece. As of now this is the only published report on Mercury 2.5, so the speed claim, its ranking, and the open question about benchmark quality all rest on that one source.