Inception, the Redwood City AI lab, released Mercury 3 on Monday. It is the newest version of what the company calls a diffusion LLM, a design that differs from the systems behind most chatbots.
Most language models are autoregressive. They write one token at a time, each one chosen after the last. A diffusion model works more like the image generators that share the name. It starts with rough, noisy output and refines many tokens at once, in parallel, until the text resolves. Inception says that parallel approach is why its model runs faster.
On the company’s own benchmarks, Mercury 3 beats OpenAI’s GPT-6 Luna, Anthropic’s Claude Haiku 5.5 and Zhipu’s GLM-5.3 Flash across a range of tests, while running four to eight times faster. Inception says it generates more than 1,000 tokens a second on Nvidia chips. Those figures are self-reported. No outside group has published a test of the model yet.
The lab says Mercury 3 is more than 40 times the size of its first model and about three times as capable across standard benchmarks, at the same speed. Its context window, the amount of text it can handle at once, has grown to one million tokens, up from 260,000 in the previous version.
Speed first, price second
Mercury 3 costs $0.20 per million input tokens and $0.70 per million output tokens. For launch, the company is halving both, to $0.10 and $0.35. Inception has not said how long that promotion will last.
Speed, not price, is the pitch. A model that answers faster is easier to run at scale, which matters most for agents and coding tools that call a model many times in a row.
“It’s comparable to the Luna and Haiku models, but it’s way faster,” Inception chief executive Stefano Ermon told The Deep View. “It’s getting 1,000 tokens per second. It’s almost instant in the way you get answers from these models.”
The launch price undercuts those rivals on output tokens, and extends a wider run of price cuts across the industry.
Built with its own agents
Inception also described how it built the model. Ermon told The Deep View that the company used swarms of Mercury 3 agents to improve the model itself, a form of recursive self-improvement. He said the technique added about 16% to the model’s speed, on top of the gains his team made by hand. Inception would not say how many parameters the model has in total, the outlet reported.
Diffusion is still a minority bet. Google DeepMind released a diffusion text model, DiffusionGemma, in June, a sign that larger labs are testing the same idea. Most of the field remains autoregressive. Independent testing will show whether the speed claims hold up outside Inception’s own benchmarks.
The model is available now through Inception’s own API, plus OpenRouter, Baseten and Vercel’s AI Gateway. It uses the same API format as OpenAI’s, so developers can switch existing code with little change. Inception’s backers include Menlo Ventures, Microsoft’s M12, Nvidia’s venture arm and the investors Andrew Ng, Andrej Karpathy and Eric Schmidt.