Inception Releases Mercury 3 Diffusion AI Model

Inception Labs has released Mercury 3, the latest model in its diffusion language architecture line. Built with over 40 times the total parameters of Mercury 1 and a threefold improvement in aggregate benchmark intelligence, the model generates text at 1,068 tokens per second on NVIDIA GPUs. It expands its context window to 1 million tokens, up from 260,000 in Mercury 2.5. Standard API rates are $0.20 per million input tokens and $0.70 per million output tokens, while launch promotional pricing drops rates to $0.10 for input, $0.35 for output, and $0.01 for cached inputs. Inception offers 100 million free tokens for new accounts, with integration available via OpenRouter, Vercel AI Gateway, Baseten, and Mercury Chat.

Unlike autoregressive models, Mercury 3 relies on parallel diffusion processing alongside features like reasoning controls, tool calling, and structured output capabilities. In inception-published evaluations, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 across benchmarks including SWE-Bench Atlas, DeepSearchQA, and GPQA Diamond. On BrowseComp+, Mercury 3 scored 55.18 compared to 50.50 for Luna and 45.78 for Haiku 5.5. Though Haiku retained leads on DeepSWE v1 and MRCR v2, and Luna won one long-context test, Mercury 3 also completed 50% of 50 live Mind2Web browser tasks, outperforming GPT-6.1 Sol.

For developers, the release opens up ultra-fast, affordable execution for multi-step agent workflows. In preliminary API testing totaling $0.04160 across eight runs, the model generated functional HTML game code for a 25-by-25 dungeon map that passed 2,000 checks, correctly calculated business metrics resulting in $900 in daily gross profit, $756 in monthly gross savings, and $576 in net benefit, and answered a logic puzzle for $0.00014. Despite minor constraint slips—such as returning 274 words when asked for 250—it successfully navigated web searches and hallucination prompts. Alongside siblings like Mercury Voice and Mercury Decide, Mercury 3 presents a viable high-speed alternative to standard models.