Mercury 2.5 posts 770 tokens per second on Artificial Analysis, topping diffusion-LM speed charts

Inception Labs' Mercury 2.5 has landed on Artificial Analysis with a measured output speed of 770 tokens per second, a figure that places the diffusion-based model ahead of autoregressive rivals on the leaderboard's throughput column. The listing is the first published speed number for the 2.5 generation, and it extends a lead the startup has claimed since Mercury's debut. Artificial Analysis runs its measurements under standardized conditions, so the comparison against other hosted models is at least methodologically consistent.

The architecture is the reason for the gap. Mercury does not emit one token at a time the way GPT, Claude, Gemini and Llama do; it refines an entire block of text in parallel passes, a diffusion approach borrowed from image generation. Inception Labs has argued that this design can hold quality roughly level with conventional models while running several times faster on identical silicon. Most hosted frontier APIs on Artificial Analysis sit in the low hundreds of tokens per second, so 770 is not a marginal improvement over the field.

Speed is the constraint that bites hardest in production. Agent loops that fire dozens of short calls, inline code completion, and interactive chat all expose latency directly to the user, and teams routinely pay premium tiers or batch requests just to hide it. A model that sustains 770 tokens per second on real hardware would shift the cost math for those workloads. That said, throughput says nothing about reasoning depth, instruction following or behavior on long contexts.

The response to the listing split along exactly that line. Several developers asked for benchmark parity before treating the speed as a win, noting that Artificial Analysis publishes quality scores alongside speed and that Mercury's position there is the number that decides whether the throughput is usable. Inception Labs has not released independent third-party evaluations to accompany the leaderboard entry, and none surfaced in the discussion around the model page.

What happens next is a two-part test. First, whether Mercury 2.5's quality scores on the same leaderboard keep pace with its speed column, since a fast model that trails on accuracy only helps a narrow set of tasks. Second, whether a major API provider adopts diffusion inference at all, which would signal the approach is more than a benchmark curiosity. Until both questions are answered, 770 tokens per second is a real number with an unresolved asterisk.