Inception Mercury 2.5 Hits 1,107 Tokens Per Second

Inception has introduced Mercury 2.5, which it claims is the largest diffusion language model ever trained. Unlike traditional autoregressive models that generate text one token at a time, this diffusion model iteratively denoises multiple tokens in parallel. This architectural shift allows Mercury 2.5 to clock speeds of 1,107 tokens per second on standard Nvidia GPUs while supporting a 260K context window. The model represents a 40 percent intelligence boost over its predecessor, Mercury 2, placing its capabilities on par with cost-optimized frontier models like Claude Haiku 4.5, Gemini 3.5 Flash-Lite, and GPT-5.6 Luna (Low).

The model is priced at $0.20 per million input tokens and $0.75 per million output tokens, though an active 80 percent launch discount temporarily reduces those rates to $0.04 and $0.15, respectively. Cached inputs cost just $0.02 per million tokens. Developers can access Mercury 2.5 via OpenAI-compatible endpoints on the Inception API, OpenRouter, and Baseten, with new accounts receiving 10 million free tokens and a 100 million token free trial. Inception evaluated the model across several benchmarks, including Tau3Bench Telecom for agentic tool use, GPQA Diamond and SciCode for scientific reasoning, IFBench for instruction following, AA-LCR for long-context recall, and TerminalBench for coding.

For practitioners, this combination of speed and low cost makes Mercury 2.5 ideal for latency-sensitive pipelines like search, retrieval-augmented generation, and voice agents. In real-world testing, the voice agent platform OpenCall used Mercury to reduce its P99 response times from several minutes to just one second, while its P50 response times dropped below 0.2 seconds. Similarly, the coding assistant Augment Code integrated the model for context compaction, routing, and Model Context Protocol tool searches, cutting latency by 82 percent—from 150 seconds down to 27 seconds—and slashing costs by 90 percent. Alongside the main release, Inception previewed Mercury Voice, which features a time-to-first-token under 170 milliseconds, and Mercury Router, a fast classifier that directs incoming prompts to the most efficient model.