Inception Mercury 3: 1,068 Tokens Per Second
Inception launched Mercury 3, a diffusion LLM beating GPT-6 Luna and Claude Haiku 5.5 on benchmarks, running 1,068 tokens per second with a 1M-token context.
Inception has launched Mercury 3, its fourth diffusion-based language model. The company says the model beats speed-optimized rivals like GPT-6 Luna and Claude Haiku 5.5 on benchmark tests while generating 1,068 tokens per second. If the numbers hold up, it breaks a trade-off developers have lived with for years: fast models or smart models, but not both.
Speed without giving up quality
Every major AI model today writes text one word at a time. Inception's diffusion models work differently: they draft a whole answer at once and refine it in parallel, like turning static into a photo. Mercury 3 is more than 40 times larger than Mercury 1, yet it runs at about the same speed, over 1,000 tokens per second on NVIDIA GPUs. The company claims Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks, running four to eight times faster. It also expands the context window to one million tokens, up from 260,000 in Mercury 2.5. That means it can read very long documents, codebases, or conversations in a single request. "Since we launched Mercury 1, the question we've heard most is whether diffusion can scale," said Stefano Ermon, Inception's co-founder. "Mercury 3 is over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second."
Who is this model for?
Inception is aiming Mercury 3 at apps where waiting hurts: browser use agents, real-time search agents, interactive tutors, customer support chats, and voice bots. Early customers are already running it in those jobs. The model is available through Inception's own API platform, plus OpenRouter, Baseten, and Vercel, with an OpenAI-compatible API. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, but a 50% launch discount cuts that to $0.10 and $0.35. New API keys also come with 100 million free tokens.
What the numbers do not prove yet
The benchmark wins are self-reported by Inception, so treat them with care. On its own benchmark table, Mercury 3 leads on 8 of 12 tests but trails Claude Haiku 5.5 on some, like DeepSWE coding tests. The 1,068 tokens-per-second figure was measured on NVIDIA GPUs the company did not name, and real-world speed also depends on network delay, prompts, and tools. Independent eyes are starting to appear. TinyFish AI reported Mercury 3 completing half of 50 live browser tasks, ahead of GPT-6.1-Sol, and Octen AI measured a median end-to-end response of 2.2 seconds. Interesting, but not yet a full verdict. The honest test will be developers putting it in production over the next few weeks.