Overview
- Inception publicly released Mercury 3 on Monday, October 19, 2026, and made it available through its API and partner platforms with introductory pricing.
- The company reports Mercury 3 runs at about 1,068 tokens per second on NVIDIA GPUs and supports a 1 million-token context window to handle much larger single requests.
- Inception says Mercury 3 is more than 40 times larger than its first model and outperforms speed-optimized rivals like GPT-6 Luna and Claude Haiku 5.5 on its internal benchmarks while costing less per token.
- Early-access partners report faster response times in production for voice, real-time tutoring, search agents and customer support, but most performance evidence so far comes from vendor-run tests and partner case studies rather than broad independent benchmarks.
- Diffusion LLMs generate and refine many tokens in parallel rather than predicting one token at a time, which explains the claimed speed gains and could reduce latency in interactive apps though wider third-party testing will be needed to confirm real-world effects.