Mercury 2.5 Preview

Inception Labs released Mercury 2.5 Preview on August 31, 2026, the latest in its Mercury line of diffusion language models, a design that generates and refines multiple tokens in parallel rather than one at a time. Inception says the approach pushes throughput to 1,107 tokens per second on standard GPUs. Inception puts the intelligence gain over Mercury 2 at around 40 percent, in range with cost-optimized models like Gemini 3.5 Flash-Lite and Claude Haiku 4.5.

Preview pricing runs at $0.04 per million input tokens and $0.15 per million output tokens, an 80 percent discount off the eventual standard rate of $0.20 and $0.75. The model supports tunable reasoning effort, parallel tool calls, and schema-aligned JSON output, and is available through Inception’s own API as well as Baseten and OpenRouter.