Inception launches Mercury 3 with claims of faster AI performance
The company reports speeds four to eight times faster than rival models in its own benchmarks, but no independent tests have been published.
Redwood City AI lab Inception has released Mercury 3, saying its latest language model outperforms several competitors while delivering faster responses. Its internal testing puts the model ahead of OpenAI’s GPT-6 Luna, Anthropic’s Claude Haiku 5.5 and Zhipu’s GLM-5.3 Flash across multiple tests.
Inception reports a speed advantage of four to eight times and output exceeding 1,000 tokens per second using Nvidia chips. The performance figures come from the company; no outside organization has yet published testing of Mercury 3.
Unlike autoregressive models, which produce tokens sequentially, Mercury 3 uses a diffusion approach that refines multiple tokens simultaneously from an initially noisy output. Inception attributes its speed to that parallel processing. The company says the model can handle one million tokens at once, compared with 260,000 for its predecessor.
Standard pricing is US$0.20 for each million input tokens, with output priced at US$0.70 per million tokens. A launch promotion reduces those rates to US$0.10 and US$0.35, respectively. Inception has not specified when the discounted pricing will end.
Chief executive Stefano Ermon told The Deep View that Inception used groups of Mercury 3 agents to help improve the model itself. He said that work increased speed by about 16%, beyond improvements made manually by his team.
Developers can access Mercury 3 through Inception’s API, OpenRouter, Baseten and Vercel’s AI Gateway. Its API follows OpenAI’s format, allowing existing code to be adapted with limited changes.
