Inception Mercury 2.5 Preview on OpenRouter Generates Tokens in Parallel at $0.15/M

What shipped

Inception Labs' Mercury 2.5 Preview — a diffusion large language model (dLLM) — went live on OpenRouter as inception/mercury-2.5-preview. Unlike every other transformer model on the platform, it generates tokens in parallel rather than one at a time.

What changed

Mercury 2.5 is a diffusion LLM: instead of predicting the next token, it produces and refines multiple tokens simultaneously, then iterates toward a coherent output. It carries a 260K token context window and supports 65K max completion tokens. Pricing is $0.04 per million prompt tokens and $0.15 per million completion tokens, with cached input reads at $0.004 per million. It supports reasoning (effort levels: high, medium, low, none), structured outputs, tool calls, and temperature/stop/top-p controls.

Why a builder cares

This is a fundamentally different architecture than GPT, Claude, or Gemini. Parallel token generation means Mercury can produce long outputs faster than sequential models at the same parameter count — and at $0.15/M out, it's competitive with frontier models on price. For agent loops, code generation, and any task where latency from long generation matters, this is the first real architectural alternative available through a standard API.