Mercury 3
Mercury 3 is a diffusion-based language model from Inception, and the lab's most capable model to date.
Specifications
| Lab | Inception |
|---|---|
| Context window | 1M tokens |
| Input price | $0.10/1M |
| Output price | $0.35/1M |
| Release | Oct 2026 |
Strengths
- Diffusion architecture generating 1,000+ tokens/sec — 4-8x faster than speed-optimized frontier rivals
- Matches or beats GPT-6 Luna and Claude Haiku 5.5 on company-reported benchmarks
- 1M-token context window, up from 260K on Mercury 2.5
- Adjustable reasoning effort (high/medium/low/none) with tool calling and structured outputs
- OpenAI-compatible API, available via OpenRouter, Baseten, and Vercel
Weaknesses
- Published benchmarks are company-reported; no independent third-party results yet
- Diffusion out-of-order generation can be harder to follow when relying on visible reasoning
- New architecture family with less ecosystem tooling than autoregressive models
Best for
- Latency-sensitive agentic workflows (browser use, search agents, multi-step tool loops)
- Real-time chat and voice support
- High-throughput data engineering and tutoring
In Depth: Mercury 3
Summary
Mercury 3 is an AI model from Inception.
Released in Oct 2026. It takes text input and produces text output, with a context window of 1M tokens. Input pricing is $0.10 per 1M tokens and output is $0.35 per 1M tokens on OpenRouter.