Inception releases Mercury 3, a model with 1,068 tokens a second

Inception released Mercury 3, a diffusion language model. A diffusion model makes many parts of a text at the same time. Inception writes that it makes 1,068 tokens each second on NVIDIA GPUs. A token is a small part of a word. Inception also writes that it is 4 to 8 times as fast as GPT-6 Luna and Claude Haiku 5.5.

What is new

Mercury 3 is more than 40 times as large as Mercury 1. It reads 1,000,000 tokens of context at one time. The price is $0.20 for each 1,000,000 input tokens and $0.70 for each 1,000,000 output tokens. At launch, the price is 50% of that.

The API is compatible with the API of OpenAI. Mercury 3 can replace a model that a system uses at this time.

The proof

Inception gave no numbers for the benchmarks. It writes that the quality is equal to GPT-6 Luna and Claude Haiku 5.5, or above it. TinyFish did a test of 50 live web tasks. Mercury 3 completed 50% of them, more than GPT-6.1-Sol.

What is not known

No other group did a test of these results. The other partners give no numbers.

Insight: Fast models for agent steps are a contest

Inception writes that an agent waits on the model at each step. It also has Mercury Decide, a model that gives a decision for a large number of items. In October, Microsoft released Microsoft-Decision-1 and wrote that it is 35 times as fast as GPT-6 Sol. This reads as a move by Inception and Microsoft to control the small, fast steps between the large models. Leaders can test these models on their agent steps.