On October 19, 2026, Inception announced Mercury 3 in a Business Wire release, the fourth release in its diffusion large language model family. Mercury 3 sustains generation speeds above 1,000 tokens per second on widely available NVIDIA GPUs, expands the context window from 260,000 tokens in Mercury 2.5 to 1 million tokens, and is more than 40 times the size of Mercury 1. Inception says Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 across a range of standard benchmarks while running 4 to 8 times faster; those figures come from the vendor's own launch materials and have not yet been independently verified.
Why diffusion generation is so fast
Conventional large models generate one token at a time in sequence. The Mercury family instead generates multiple tokens in parallel through a diffusion process and refines them iteratively. The route's historical weakness was scale: Mercury 1, launched in June 2025, handled simple tasks such as code completion, Mercury 2 added reasoning, and Mercury 2.5 added tunable reasoning depth. Mercury 3 is the first release on this route to combine scale, a long context window and speed in one product. Co-founder Stefano Ermon said in the announcement that the limits on scaling diffusion models are mainly compute and data rather than fundamental algorithmic barriers, and that training of the next model has already begun.
Pricing and access
Mercury 3 is priced at $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount bringing that to $0.10 and $0.35, and cached tokens billed at 90% off the input price. It is available now through the Inception API, Baseten, OpenRouter and Vercel, with an OpenAI-compatible API, so existing projects can switch by changing a model name. Early customers are already using it in browser agents, interactive education, real-time search agents and customer support; education platform RevisionDojo reports that hints and marked answers come back in about a second, at quality as good as or better than the models it ran before. Teams chasing output speed can compare it with the MiniMax M3.1 Flash covered earlier on this site: one relies on parallel architecture, the other on engineering optimization, at a similar price point.
Speed matters most in agent loops: the model waits on itself at every step, so latency saved per step multiplies across steps. Teams considering a trial should check three things first: whether their workload is actually latency-sensitive; how the vendor benchmarks relate to their own tasks, ideally by testing with real business samples before migrating; and that the half price is a launch offer, so long-term costs should be estimated at the standard rate.
