Mercury 3

Not yet independently verified. Rests on Inception's own announcement; the speed of 1,068 tokens per second is Inception's measurement, and early users reported lower speeds when calling the API directly. The parameter count and licence are not stated. We will update this when it can be confirmed, and remove this note.

Mercury 3 is a diffusion-based language model from Inception, released on 19 October 2026 as the flagship of its Mercury line. Instead of writing one token at a time, a diffusion model refines a whole block of text over several passes, which is where Inception's speed claims come from.

What it does

Inception says Mercury 3 is more than 40 times larger than Mercury 1 in total parameters and runs at 1,068 tokens per second on NVIDIA GPUs, and that it matches or beats speed-optimised frontier models on a range of benchmarks; those comparisons are the company's own, so no score is stored here. The context window is 1 million tokens, up from 260,000 for Mercury 2.5. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a launch offer at half that, and cached input tokens are 90% off the input price.

Who it is for

The model is meant for latency-sensitive work where an application waits on the model at every step: browser and web-navigation agents, real-time search, voice and chat customer support, and interactive tutoring. Inception names early-access users including TinyFish, RevisionDojo, Octen and Lorikeet, and sells Mercury Voice, a variant tuned for voice agents, alongside it. Users who tested it after launch reported speeds well below the stated figure when calling the API directly, so throughput is worth measuring on your own workload.

Key facts

Maker: Inception Released: 19 October 2026 Context: 1M tokens Price: $0.20 input, $0.70 output per 1M tokens (launch offer $0.10 and $0.35) Architecture: diffusion language model Weights: not released