TL;DR
Inception launched Mercury 3 on October 19. The company reports 1,068 tokens per second on NVIDIA GPUs and benchmark results competitive with GPT-6 Luna and Claude Haiku 5.5. Its listed specifications include a 1-million-token context window and launch token rates at half the standard price. The performance figures are vendor-reported; this briefing has not independently measured the model.
What Mercury 3 is
Inception announced Mercury 3 on Monday, October 19, calling it its most capable diffusion-based large language model. Diffusion LLMs generate and refine multiple tokens in parallel rather than emitting every token sequentially. Inception reports 1,068 tokens per second on NVIDIA GPUs and says Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running four to eight times faster. These are Inception's comparisons, not independent measurements by this briefing or guarantees for every workload.
Inception lists a 1-million-token context window, up from 260,000 for Mercury 2.5. Its model page advertises launch rates of $0.10 per million input tokens, $0.35 per million output tokens, and $0.01 per million cached input tokens; standard listed rates are $0.20, $0.70, and $0.02 respectively. The pages checked do not specify an end date for the launch discount. The Deep View reports that co-founder Stefano Ermon hopes further developments could make the lower price permanent; that is a possibility, not a pricing commitment.
The scaling question
Inception frames Mercury 3 as evidence that text diffusion can scale. It says total parameters have increased more than 40-fold from Mercury 1, and describes a threefold rise in intelligence across a range of standard benchmarks from Mercury 1 (mini) to Mercury 3 while reported speed remains above 1,000 tokens per second. That is the company's characterization of benchmark progress, not proof that the model is three times as capable on every task.
The announcement does not disclose Mercury 3's absolute parameter count. Relative model-size and benchmark comparisons are useful context, but builders still need to evaluate the service on the tasks and serving conditions they intend to use.
Where you can get it, and what's going away
Mercury 3 is available now through the Inception API Platform, OpenRouter, Baseten, and Vercel, and it speaks the OpenAI API format — the model id is `mercury-3`, with `reasoning_effort` and `temperature` parameters and official Python and TypeScript SDKs. New platform accounts get 100 million free tokens with no payment details required.
Alongside the launch, Inception is retiring the older line: Mercury 2 and 2.5 are no longer available to new users. Existing customers with usage before October 19 keep access until November 15, 2026, and enterprise customers stay supported under their existing agreements.
What matters for builders
Our view: Mercury 3 is worth evaluating for latency-sensitive agents and interactive applications. Inception's launch page quotes RevisionDojo reporting feedback in about a second, and Octen AI reporting median generation time of 0.9 seconds with median end-to-end response time of 2.2 seconds in its tests. These are partner testimonials presented by Inception, not results reproduced for this briefing. Separately, The Deep View reports Inception's claim that Mercury 3 agents helped improve model speed by 16%. Builders should measure their own latency, completion rate, and cost per completed task.
Inception also advertises Mercury Voice with time to first token under 170 milliseconds and says Mercury Decide outperforms Jev on benchmarks including JevBench and LegalBench. Those are vendor claims about separate products, not additional Mercury 3 results. Whether the launch discount becomes permanent and whether future models retain the advertised speed remain open questions.