Inception Labs released Mercury 3 on October 19 and led with a speed. One thousand and sixty-eight tokens a second, on Nvidia GPUs, from a model that writes by cleaning up noise instead of marching one word at a time.
What the company is claiming
Diffusion language models guess a whole span and then refine it. Inception says that is why Mercury 3 can post a throughput number in the thousands while still sitting, on its own charts, next to GPT-6 Luna and Claude Haiku 5.5. It also says the new model has more than 40 times the parameters of Mercury 1, which is their way of arguing that this architecture scales. Context is a million tokens.
There is one agent anecdote in the launch notes. Early-access partner TinyFish said Mercury 3 finished half of 50 live web tasks in a Mind2Web-style set. Fifty tasks is a pilot, not a census. Inception is quoting its partner.
Price, and the reason to wait a week
The rate card at launch is $0.20 per million input tokens and $0.70 per million output tokens. Inception says those numbers are halved for the launch, so do not bake them into a yearly forecast without checking the date.
Speed this aggressive is worth a test on your own prompts. It is not worth migrating a product because a vendor’s chart beat a small model from someone else. Ask which benchmark, which settings, and whether the fast mode is the accurate mode.