Every fast AI model you have ever used writes one token at a time, and a long answer takes a long time because word six cannot start until word five exists. On October 19, Inception Labs shipped a model that refuses to play that game. Mercury 3 is a diffusion language model: it drafts an entire answer as noise and refines the whole thing in parallel, the way an image generator resolves static into a picture.
I wrote up the full guide on ToolPilot Pro, with the spec sheet, the pricing math, what Inception's benchmark chart does and does not prove, and the four catches under the headline:
Read the full guide: Mercury 3 vs GPT-6 Luna
The numbers that matter
- 1,068 tokens/sec on NVIDIA GPUs (Inception's own measurement), a claimed 4 to 8 times faster than speed-tuned rivals
- 1M token context, up from 260K in Mercury 2.5
- $0.20 / $0.70 per million input/output at list, halved for launch to $0.10 / $0.35, cached input 90% off
- On Inception's chart it leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks, including BrowseComp+ (55.18 vs 50.50 vs 45.78); it trails Haiku on DeepSWE v1 and MRCR v2
- API-only, no open weights: Inception API, OpenRouter, Baseten, Vercel, all OpenAI-compatible
The honest caveat
Every speed and benchmark figure here is Inception's own. Independent verification has not happened yet, and the launch price is a promotion, not a permanent rate. The sensible move is a 50-prompt test on your own workload, which at these prices costs less than a coffee.
Full guide: https://toolpilotpro.com/mercury-3-diffusion-guide
Top comments (0)