# Mercury 3, Explained

> Mercury 3 is Inception's fourth diffusion language model (announced October 19, 2026). It generates and refines many tokens in parallel instead of writing word by word. Inception reports 1,068 tokens/sec on NVIDIA GPUs and a 1M-token context window. This page includes an animated illustration of diffusion vs autoregressive generation, an API cost calculator vs Claude Haiku 5.5 and GPT-6 Luna, a fact table, model timeline, and FAQ. All benchmark and speed figures are company-reported, not independently verified (checked October 20, 2026).

Mercury 3 is Inception's newest diffusion LLM (dLLM), announced October 19, 2026 via Business Wire. Unlike standard autoregressive models, diffusion models draft a full answer of masked tokens, then refine every position in parallel over a few passes. Inception calls Mercury 3 the most capable diffusion LLM ever released, claiming quality on par or better than GPT-6 Luna and Claude Haiku 5.5 across standard benchmarks while running 4-8x faster.

Key facts (all company-reported):
- Speed: 1,068 tokens/sec on widely available NVIDIA GPUs
- Context window: 1,000,000 tokens (Mercury 2.5 had 260,000)
- Scale: more than 40x the parameters of Mercury 1
- List price: $0.20 per 1M input tokens, $0.70 per 1M output tokens; cached input 90% off ($0.02)
- Launch price: 50% off list — $0.10 input, $0.35 output, $0.01 cached input; no end date announced
- Launch output price is about 30% below GPT-6 Luna and Claude Haiku 5.5 short-tier output ($0.50)
- Access: Inception API, OpenRouter, Baseten, Vercel; OpenAI API compatible; API only, no open weights
- Mercury 2 / 2.5: closed to new users; non-enterprise access ends November 15, 2026
- CEO Stefano Ermon says the company is trying to make launch pricing permanent

Model history: Mercury 1 (June 2025, first commercial diffusion LLM) -> Mercury 2 (reasoning) -> Mercury 2.5 (tunable reasoning, 260K context) -> Mercury 3 (October 2026, biggest intelligence jump, 1M context).

Company-reported customer results: RevisionDojo (~1s tutor feedback, drop-in swap), Octen AI (0.9s median generation, 2.2s end-to-end, ~80ms search), TinyFish AI (50% of 50 live Mind2Web navigation tasks).

Guide: https://lab.awaara.cn/mercury-3/
Official announcement: https://www.businesswire.com/news/home/20261019090400/en/Inception-Launches-Mercury-3-Setting-a-New-Pareto-Frontier-in-Speed-and-Intelligence
Related: GPT-6 Luna pricing https://lab.awaara.cn/gpt-6-luna-pricing/ , Claude Haiku 5.5 https://lab.awaara.cn/claude-haiku-5-5/