Mercury 3

Launched October 19, 2026 by Inception Labs, Mercury 3 is a diffusion-based large language model (dLLM): instead of generating one token at a time like conventional LLMs, it drafts the whole answer and refines it in parallel. Inception claims 1,068 tokens per second on NVIDIA GPUs — 4–8x faster than rivals — with sub-300ms time to first token and up to 70% lower cost per task. It ships with a 1 million-token context window, up from 260K in Mercury 2.5. Note: the speed and benchmark figures are company-reported; Inception says quality is on par with or better than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on selected benchmarks (IFBench, GPQA Diamond, SWE-Bench Atlas), but these have not been independently verified.

Key facts: API-only access (no open weights) through an OpenAI-compatible API, available on Inception's platform plus OpenRouter, Baseten and Vercel. A Mercury Voice variant is tuned for voice agents with time-to-first-token under 170ms, and Mercury Decide is a decision model for structured choices. Founders of Inception pioneered diffusion modeling and invented cornerstone techniques behind modern image and video generation.

Plans: genuine free access — 100M free tokens for new API keys. Paid list pricing: $0.20 input / $0.70 output per 1M tokens, with a 50% launch discount currently showing $0.10 / $0.35 per 1M; cached input is 90% off.

Best for: developers building real-time agents, voice bots, browser agents, customer support and search experiences where latency is the product.

Related: Claude · Cursor · ChatGPT

Pros

• Diffusion-based generation drafts the whole answer and refines it in parallel, instead of one token at a time.
• Company-reported speed: 1,068 tokens per second on NVIDIA GPUs, 4 to 8 times faster than rivals, with sub-300ms time to first token and up to 70% lower cost per task.
• A 1 million-token context window, up from 260K in Mercury 2.5.
• API-only access through an OpenAI-compatible API, available on Inception's platform plus OpenRouter, Baseten and Vercel, with a Voice variant for voice agents and a Decide model for structured choices.

Watch out for

• The speed and benchmark figures are company-reported and have not been independently verified, per the page's own note.
• API-only access with no open weights, so you cannot self-host the model.
• Launched October 19, 2026, so the ecosystem around it is still young.

Best for

Developers building real-time agents, voice bots, browser agents, customer support and search experiences where latency is the product. Who should skip it: users looking for a chat app or for open weights to self-host.

Pricing

Genuine free access with 100M free tokens for new API keys. Paid list pricing is $0.20 input and $0.70 output per 1M tokens, with a 50% launch discount currently showing $0.10 and $0.35 per 1M; cached input is 90% off. Check the official pricing page for current pricing.

FAQs

What is a diffusion LLM? Instead of generating one token at a time like conventional LLMs, it drafts the whole answer and refines it in parallel.

How fast is Mercury 3? Inception claims 1,068 tokens per second on NVIDIA GPUs with sub-300ms time to first token; these are company-reported figures, not independently verified.

How do I access it? API-only (no open weights) through an OpenAI-compatible API, on Inception's platform plus OpenRouter, Baseten and Vercel.

What are Mercury Voice and Mercury Decide? A Voice variant tuned for voice agents with time-to-first-token under 170ms, and a Decide model for structured choices.