Inception's new diffusion LLM runs at over 1,000 tokens per second and expands the context window to 1 million tokens, with a 50%-off launch price.
2026-10-19 · by Moe Ameen
On October 19, 2026, Inception (Inception Labs) released Mercury 3, which it calls the most capable diffusion-based large language model released to date. Where a standard autoregressive model writes one token at a time from left to right, a diffusion LLM (dLLM) starts from a rough draft of the whole output and refines tokens in parallel — the mechanism behind its speed. Inception reports 1,068 tokens per second on NVIDIA GPUs and says Mercury 3 runs 4–8x faster than GPT-6 Luna and Claude Haiku 5.5 while matching or beating them on quality.
The headline changes over Mercury 2.5 are scale and context. Inception says Mercury 3 is more than 40x larger than Mercury 1 in total parameters, and the context window jumps to 1 million tokens, up from 260K in Mercury 2.5. It positions the model in the frontier speed-optimized tier, describing quality as on par with or better than other fast frontier models at a fraction of the latency. Specific benchmark names and scores were not published alongside the launch.
On pricing and access: standard rates are $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount to $0.10 input and $0.35 output, and cached input tokens 90% off. Mercury 3 is available through Inception's API platform, OpenRouter, Baseten, and Vercel over an OpenAI-compatible endpoint, with 100 million free tokens to evaluate it. Early-access partners in browser agents, education, search, and customer support reported latency and eval wins in testing, though those figures are company-reported rather than independently verified.
The practical move this launch week is to treat Mercury 3's pricing as a near-free drafting budget. With 100 million free tokens and a 50%-off promo, over-generate without worrying about cost: run a transcript or a batch of ideas through it and pull back fifty hooks, ten script angles, or a week of caption options for almost nothing. The catch is the one every raw model leaves you with — a pile of fast drafts is not a content calendar. That hand-off is exactly where [Kompozy](/) picks up. Paste a winning Mercury 3 draft into Kompozy and it becomes a captioned [Persona Short](/glossary/persona-shorts) (a HeyGen avatar reading your script), a brand-exact [Carousel](/glossary/hyperframes), Quote Graphics, and per-platform text posts, each governed by a [Persona Brief](/glossary/persona-brief) so the fast copy still lands in your voice.
Then [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email, each piece passing a per-post review step first. Mercury 3 made the words cheap; Kompozy is the layer that turns them into finished, on-brand content and puts it everywhere. For the evergreen tool page on this model, see [Mercury 3 in AI Tools](/ai-tools/mercury-3); for the previous release, [Mercury 2.5](/ai-tools/mercury-2-5).
Mercury 3 is a diffusion large language model (dLLM) from Inception, released October 19, 2026. Instead of generating one token at a time, it refines tokens in parallel, which Inception reports at 1,068 tokens per second — 4–8x faster than comparable fast frontier models. It also raises the context window to 1 million tokens.
Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount to $0.10 input and $0.35 output, and cached input tokens 90% off. There are 100 million free tokens to evaluate it. Launch promotions expire, so budget against the standard rates.
Mercury 3 raises the context window to 1 million tokens (from 260K), scales to more than 40x the parameters of Mercury 1, and Inception reports higher throughput and quality on par with or better than other fast frontier models. Both are diffusion LLMs that refine tokens in parallel for speed.
No. Mercury 3 is a text model and API — it drafts words and stops. It has no image or video generation, no brand-voice governance, and no scheduling or publishing. To turn its drafts into finished, posted content you pair it with a content engine; Kompozy, for example, brands the copy, generates the media, and publishes across platforms.