US AI company Inception on October 19 released Mercury 3, a diffusion-based large language model that it calls the most capable diffusion LLM to date. The model is live on Inception's own API platform and is also available through OpenRouter, Baseten and Vercel.
The company gave several headline figures: generation speed of 1,068 tokens per second on Nvidia GPUs; a context window of 1 million tokens, versus 260,000 for the previous Mercury 2.5; and a total parameter count more than 40 times that of the original Mercury 1. The exact number of parameters has not been disclosed.
Speed and pricing
Inception says Mercury 3 beats OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 across a range of benchmarks while running 4 to 8 times faster than those comparison models. The blog post, however, does not list per-benchmark scores. It says only that, on standard benchmarks, the model is roughly 3 times as intelligent as Mercury 1 (mini). Independent evaluations from outside organizations have not been published yet.
Two tests come with concrete numbers. The first is a browser agent task: on 50 real Mind2Web web-navigation tasks, Mercury 3 completed half of them, which the company says is “significantly more than GPT-6.1-Sol.” The second comes from partner Octen AI, whose Sense API uses Mercury 3 for generation: median generation time is 0.9 seconds and median end-to-end response time is 2.2 seconds, of which Octen's own search accounts for only about 80 milliseconds.
Pricing is $0.20 per million input tokens and $0.70 per million output tokens. During the launch period it is half price, at $0.10 and $0.35, and cached input tokens get an additional discount. New users of the API platform receive 100 million free tokens.
In its blog post, Inception frames the significance of the release in terms of scaling:
“The scaling laws that drove progress in autoregressive models hold for diffusion too.”
Who else is pursuing diffusion this year
Mainstream large models write text autoregressively: they emit one token after another, and each token has to wait for the one before it. Diffusion models borrow from image generation. They first lay down a rough draft of an entire passage and then refine it over several denoising passes, updating many positions at once. That is where the speed advantage comes from.
The approach has clearly heated up this year. In May, Nvidia's Nemotron Labs released Nemotron-Labs-Diffusion in 3B, 8B and 14B sizes, where a single model can switch between autoregressive and several diffusion decoding modes. In June, Google open-sourced DiffusionGemma, which generates 256 tokens in parallel in one step. At this year's ICML, a top award also went to a diffusion paper.
The strategies differ from one player to the next. Nvidia and Google have released small and mid-sized open models aimed at research and on-device use. Inception has stayed with a closed, commercial API, using speed and price to go after latency-sensitive workloads such as code completion, voice assistants, and agent workflows that call a model dozens of times in a row. With Mercury 3, it has raised parameter count and context length together, trying to show that diffusion models keep their capability as they scale up, and are not just “fast but dumb.”
Whether that holds depends on two things: scores on independent leaderboards, and whether the platforms that have wired it into production are still using it a few months from now. The only customer figures in the official blog post come from Octen, and the sample is still too small.
The previous generation, Mercury 2, was named one of TIME's Best Inventions of 2026 just a few days ago. Public reports had earlier said that both Microsoft and SpaceX were in talks with Inception; there is no new public information on the status of any deal.
Sources: Inception official blog (for speed, context window, Mind2Web completion rate, standard API pricing and launch discount), OpenRouter model page, CocoLoop, technical documentation on diffusion models from Google and Nvidia.