Inception launched Mercury 3 on October 19, 2026, positioning the diffusion-based large language model as a faster, cheaper alternative to autoregressive frontier models2,3. The company claims Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running four to eight times faster, generating 1,068 tokens per second on NVIDIA GPUs1.
List pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount bringing effective rates to $0.10 per million input tokens and $0.35 per million output tokens. Cached input tokens are priced at $0.02 per million at list rate and $0.01 under the promotion. The context window reaches one million tokens, up from 260,000 tokens in Mercury 2.5.
Inception framed Mercury 3 as evidence that text diffusion scales competitively. The company said Mercury 3 is more than 40 times larger than Mercury 1 in total parameters, while intelligence across standard benchmarks rose threefold from Mercury 1 (mini) to Mercury 3. Inception described Mercury 1 as the first proof point that diffusion LLMs could be built for commercial applications and said the open question had been whether diffusion LLMs could scale to compete with autoregressive models on complex use cases.
CEO and co-founder Stefano Ermon said the model delivers more than 1,000 tokens per second while improving on the company's earlier models.
All published benchmark results and speed figures are company-reported. Mercury 3 enters a frontier landscape that has shifted rapidly: OpenAI began rolling out GPT-6 to all ChatGPT users on October 8[1], and Mistral released the 1.05-trillion-parameter Mistral Large 4 under open weights on October 7[3]. ANALYSIS Mercury 3's competitive claims rest on comparisons to Luna and Haiku rather than to the newest autoregressive releases, and independent reproduction of the benchmarks will determine whether diffusion-based generation can hold its speed advantage at frontier quality levels.