Inception Mercury 3 Review: Diffusion LLM Speed, Pricing and Use Cases

Inception Mercury 3 reached general availability on October 19, 2026, as the latest diffusion-based LLM from Inception Labs. The model produces text by refining blocks of tokens in parallel instead of generating them sequentially. This approach yields reported output speeds of 1,068 tokens per second on NVIDIA GPUs while maintaining a 1-million-token context window.

Inception Mercury 3 Overview and Key Capabilities

The model supports tool calling, structured outputs, and adjustable reasoning effort levels labeled instant, low, medium, and high. Maximum output length reaches approximately 131,000 tokens. Parameter count grew more than 40 times compared with Mercury 1, and the company reports roughly three times higher scores on standard benchmarks than the prior version. An OpenAI-compatible API endpoint allows integration through Inception Labs, OpenRouter, Baseten, and Vercel.

Early partner deployments highlight latency-sensitive applications. TinyFish reported 50 percent success on 50 Mind2Web browser tasks at costs comparable to GPT-6.1-Sol. RevisionDojo measured roughly one-second feedback loops in tutoring scenarios. Octen recorded median generation times of 0.9 seconds for smart-glasses search queries. These cases show the speed advantage compounds in agent, voice, and real-time search workloads.

Pricing, Benchmarks, and Practical Limitations

Standard API rates list at $0.20 per million input tokens and $0.70 per million output tokens. A launch promotion halves those figures to $0.10 input and $0.35 output. Cached input tokens receive an additional 90 percent discount to $0.01. New accounts receive 100 million free tokens upon signup. No duration for the promotional rates has been announced.

Independent test suites on OpenRouter show strong performance on short reasoning and coding tasks, with one 306-test run reaching 93 percent accuracy at roughly half the median time of comparable models. Long-context results remain mixed in vendor data, with trailing scores on MRCR v2 and AA-LCR. Production variability appears on complex multi-step creative tasks.

Reported limitations include reduced gains when heavy reasoning levels are enabled and inconsistent results on extended creative or long-horizon planning. No independent security audits or data-handling certifications have been published yet. Early user reports note occasional quality drops on tasks exceeding several thousand tokens.

Direct comparisons place Mercury 3 ahead on eight of twelve vendor benchmarks against GPT-6 Luna and Claude Haiku 5.5, though trailing on certain long-context evaluations. Speed remains the clearest differentiator for workloads where per-step latency matters more than peak benchmark scores.