Inception announced Mercury 3, a new diffusion-based large language model that it says generates text at around 1,000 tokens per second, far faster than mainstream chatbots. The company claims the model is comparable in quality to low- and mid-tier systems such as Luna and Haiku, but 4x to 8x faster.
The launch lands as a direct challenge to the industry's current playbook: build ever-larger LLMs and rely on more Nvidia hardware to make them usable. Inception says Mercury 3 changes the equation by generating and refining several tokens at once instead of one at a time, which makes better use of GPU parallelism.
Mercury 3 speed, price and benchmarks
The pricing is aggressive but temporary at first. Mercury 3 costs $0.10 per million input tokens and $0.35 per million output tokens, which Inception says matches Haiku 5.5 on input and undercuts it by 35% on output. That figure reflects a launch promotion at half the eventual standard price, and Inception has not said how long the discount will last.
CEO Stefano Ermon told The Deep View that the company is working on developments that could let it keep the launch price permanently in the coming weeks. Even so, Inception frames Mercury 3 primarily as a flagship speed model, not a cheap budget option like the entry-level models from Anthropic and OpenAI.
- Model type: diffusion LLM, not a traditional autoregressive LLM.
- Speed: Inception claims 4x to 8x faster than Haiku 5.5, GPT-6 Luna and GLM-5.3 Flash.
- Pricing: launch price of $0.10 per million input tokens and $0.35 per million output tokens (50% off the planned standard price).
