Inception on October 19, 2026 released Mercury 3, the newest model in its Mercury family of diffusion-based large language models. The company calls it the most capable diffusion LLM released to date and positions it as a speed-first option for agents, voice and chat support, real-time search, and tutoring apps.

Diffusion models take a different route to text than the autoregressive systems behind most chatbots. Rather than producing one token after another, they draft a full response and refine many positions in parallel over a few passes. Inception’s pitch has always been that this approach cuts waiting time; the open question was whether the quality would hold up as the models grew.

Inception says Mercury 3 answers that question. Per the company, the model has more than 40 times the total parameters of Mercury 1, and its average score across standard benchmarks is about three times higher than the original Mercury 1 mini, while generation speed stays above 1,000 tokens per second. Inception reports 1,068 tokens per second on NVIDIA GPUs, which it says is four to eight times faster than OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5. The company has not disclosed Mercury 3’s exact parameter count.

The context window grows to 1 million tokens, up from 260,000 in Mercury 2.5. List pricing is $0.20 per million input tokens and $0.70 per million output tokens; at launch Inception is charging half that, $0.10 and $0.35, with cached input discounted 90% from the input price.

On quality, Inception says Mercury 3 is on par with or better than GPT-6 Luna and Claude Haiku 5.5. According to Inception’s benchmark table, run at medium settings and summarized by MarkTechPost, Mercury 3 leads both rivals on 8 of 12 tests, including 51.08 on SWE-Bench Atlas, 76.84 on IFBench, and 89.23 on GPQA Diamond. It trails Haiku 5.5 on DeepSWE v1 and the MRCR v2 long-context test. These are Inception’s own numbers; DigiEditorial did not find independent benchmark results for Mercury 3 at the time of writing.

Inception also quotes early-access partners. Web-agent company TinyFish says Mercury 3 completed half of 50 live Mind2Web navigation tasks in its harness, and Octen AI reports a median generation time of 0.9 seconds for a smart-glasses search use case. Those are customer statements rather than independent evaluations.

Mercury 3 is available now on Inception’s API platform and through OpenRouter, Baseten, and Vercel, with an OpenAI-compatible API and 100 million free tokens for new API keys. It is API-only; Inception is not releasing the weights. The company says Mercury 2 and Mercury 2.5 are no longer available to new users, while existing customers with prior usage keep access until November 15, 2026.

DigiEditorial treats Inception’s October 19 launch post and model page as the primary sources, and corroborated the release, speed, pricing, and benchmark claims against MarkTechPost’s October 19 report, the OpenRouter model listing, and Inception’s Business Wire announcement.