Inception's diffusion model Mercury 3 tops 1,000 tokens a second

On 19 October Inception released its flagship model Mercury 3. Unlike most chat models, it does not write word by word: it drafts the whole answer and refines it step by step. Inception says it exceeds 1,000 tokens per second; one report cites 1,068 on Nvidia GPUs.

Mercury 3 has a 1-million-token context window. List price is $0.20 per million input tokens and $0.70 per million output tokens, halved at launch to $0.10 and $0.35. Inception claims it beats GPT-6 Luna and Claude Haiku 5.5 on various benchmarks while running four to eight times faster. These are company-reported figures.

Most image models use diffusion, refining noise into a picture. Inception applies the idea to text. Classic autoregressive models produce each token after the previous one, so long answers get slower; diffusion works on the whole answer at once in fewer steps. The trade-off can be quality on long, complex reasoning.

Inception, founded by researchers from Stanford and UCLA, raised $50 million in November 2025. Its 2025 Mercury report already showed over 1,000 tokens per second on H100s.

Agents call models dozens of times per task, so latency adds up. Speed now matters right after quality, especially for code completion, voice assistants and live support. MiniMax launched M3.1 Flash the same week, claiming 70% lower usage costs.

Diffusion models suit short, structured outputs needed fast: code edit suggestions, form filling, classification, voice replies and an agent deciding its next step. Long multi-step analysis still favours large autoregressive models, so models like Mercury 3 are more likely to work alongside them as fast helpers than to replace them.

• Track latency as closely as quality.
• Route tasks by model: fast ones for short answers, large ones for complex reasoning.
• Budget on list prices, not launch discounts.