Inception launches Mercury 3 diffusion language model

Inception introduced Mercury 3 on October 19, 2026, calling it the company’s most capable diffusion language model to date.[1][2] Inception reports that the model runs at 1,068 tokens per second on NVIDIA GPUs and has a one-million-token context window, up from 260,000 tokens for Mercury 2.5.[3][1] The company says Mercury 3 is more than 40 times larger than Mercury 1, and that benchmark intelligence rose threefold between Mercury 1 mini and Mercury 3.[1][3] Inception describes the release as evidence that diffusion models can scale without giving up speed.[1]

Mercury 3’s listed price is $0.20 per million input tokens and $0.70 per million output tokens.[1] Inception discounted those rates by 50% at launch, to $0.10 per million input tokens and $0.35 per million output tokens.[1] The company says cached tokens cost 90% less than the input-token price and offers 100 million free tokens for API use.[1][3] Mercury 3 is available through Inception’s API, Baseten, OpenRouter and Vercel.[1] Inception says its OpenAI API compatibility allows teams to adopt the model without rewriting code.[1] Teams assessing a switch should distinguish launch pricing from the listed rates.

Inception says early-access partners are using Mercury 3 for voice and chat support, real-time tutoring, data engineering and web navigation.[3] It also says enterprise deployments support dedicated capacity, autoscaling, compliance controls and configurable data retention.[1] Inception reports a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds.[3] These use cases suggest areas teams could evaluate in a pilot, while the cited results remain Inception’s reports.

Inception says Mercury 3 outperformed GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running four to eight times faster.[1] In its production web-agent test, the company says Mercury 3 completed half of 50 live Mind2Web navigation tasks.[3] It also reports that Mercury 3 completed significantly more tasks than GPT-6.1-Sol using the same token volume and number of agent steps, at a fraction of the per-token cost.[3] These results are not independent evaluations: BenchLM said the figures were vendor-reported and independent scores were not yet available.[4] Inception co-founder Stefano Ermon said the company had started training its next model.[1]