Inception releases Mercury 3 with faster diffusion inference and 1M-token context
The AMW Read
Reported speed, context and capability gains meaningfully update diffusion models' competitive position, while explicit scaling-law claims remain company-reported rather than debate-resolving evidence.
Inception releases Mercury 3 with faster diffusion inference and 1M-token context
Inception announced Mercury 3 on October 19, claiming its diffusion-based language model outperforms GPT-6 Luna and Claude Haiku 5.5 across multiple benchmarks while running four to eight times faster. The company reports 1,068 tokens per second on NVIDIA GPUs and a 1 million-token context window, up from Mercury 2.5's 260,000. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount cutting both prices in half.
The release positions diffusion as a competing architecture in the general-purpose model market, with latency and price as its commercial entry points. Inception says Mercury 3 has more than 40 times Mercury 1's parameters and three times Mercury 1 mini's benchmark-measured intelligence, arguing that scaling laws also apply to diffusion models. Those company-reported results strengthen the case for testing the architecture; they do not independently establish broad superiority. Availability through OpenRouter, Baseten and Vercel, alongside Inception's API, gives developers several access routes.
Builders evaluating browser agents, real-time search or customer support should compare task completion, end-to-end latency and cost per successful task against their existing models. Early-user Octen AI reports median generation time of 0.9 seconds but median total response time of 2.2 seconds, illustrating why throughput alone is insufficient. OpenAI API compatibility could ease integration. Existing Mercury 2 and 2.5 customers also face a migration deadline: access continues through November 15, with enterprise support governed by contracts.