Inception released Mercury 2.5 Preview on August 31, 2026, a diffusion large language model that generates and refines many tokens in parallel instead of one after another. The result is speed: Inception clocks it at up to 1,107 tokens per second on standard GPUs, a roughly 10-point intelligence jump over Mercury 2, at a cost-optimized price of $0.04 per million input tokens and $0.15 per million output.
What This Enables
If latency is the bottleneck in something you are building, Mercury 2.5 is worth trying today. It is live exclusively on OpenRouter as inception/mercury-2.5-preview, so you can point an existing OpenAI-compatible client at the endpoint, keep your prompts, and feel the parallel decoding as faster responses. It targets three jobs where speed compounds: search and research agents that fire many short calls, voice pipelines that need sub-second turns, and coding subagents that fan out edits across a repo. Parallel tool calls, tunable reasoning levels, and schema-aligned JSON output are built in, so structured workflows drop in without extra prompt scaffolding.
Why It Matters
Most frontier models still decode one token at a time, which sets a hard floor on how fast a response can arrive. A diffusion LLM sidesteps that floor by refining a whole block of tokens at once, and Mercury 2.5 pushes both the speed and the quality of that approach up a tier. Inception positions the quality against cost-optimized frontier models like Claude Haiku 4.5 and Gemini 3.5 Flash-Lite, so the trade you weigh is no longer speed versus intelligence but speed at a similar intelligence for a fraction of the token cost. That continues the steady drop in API prices reshaping how builders pick a default model.
Key Details
Model: Mercury 2.5 Preview, a diffusion LLM from Inception.
Context: 260,000 tokens, with up to 65,536 completion tokens.
Pricing: $0.04 per million input tokens, $0.15 per million output, and $0.004 per million cached-read tokens. Inception is running 80 percent off through September 7.
Speed: up to 1,107 tokens per second claimed on standard GPUs.
Availability: exclusive to OpenRouter during the preview. It sits in the same cost-optimized bracket you would weigh against open coding models like GLM-5.3 Flash and Qwen when choosing an agent backbone.
What to Do Next
Add the OpenRouter model id to a non-critical path first, a draft generator, an autocomplete helper, or one subagent in an existing pipeline, and measure end-to-end latency against your current model before committing. Because this is a preview with promotional pricing, treat it as an evaluation window: benchmark output quality on your own tasks now, and decide whether the speed gain justifies a switch once standard pricing lands after September 7.
