Mercury 2.5 hits 1,107 tokens/sec: price the fast path

Inception launched Mercury 2.5 on September 8, and the number that matters is 1,107 tokens per second on ordinary NVIDIA GPUs. That is not a frontier-model claim and it is not trying to be. It is a claim about the fast path — the part of your product where a user is sitting there watching a cursor blink, and every 200 milliseconds costs you a little more of their attention.

What actually happened

Mercury 2.5 is a diffusion LLM. Instead of generating one token after another, it refines a block of output in parallel, which is where the throughput comes from. Inception was founded by the Stanford, UCLA and Cornell researchers behind the first diffusion LLM, so this is the third iteration of a real line of work, not a benchmark stunt.

The specs: a 260K context window, tunable reasoning, parallel tool calls, and schema-aligned JSON output. List pricing is $0.20 per million input tokens and $0.75 per million output, with an 80% launch discount currently putting it at $0.04 and $0.15. It is available through Inception's API, OpenRouter and Baseten.

Inception positions the quality as comparable to cost-optimized frontier tiers — GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, Claude Haiku 4.5 — and claims a 40% intelligence gain over Mercury 2. Treat that framing as the vendor's, and benchmark it on your own tasks. The throughput figure is the one that is hard to fake in production.

Why it matters for your business

Most teams pick one model and route everything through it. That is a pricing mistake and a UX mistake at the same time, because the work is not uniform. Classifying a support ticket, extracting fields from an invoice, drafting an autocomplete suggestion, deciding which of four tools to call — none of that needs a reasoning model, and all of it needs to feel instant.

Split the routing. Send the interactive path — anything a human is waiting on, anything inside a voice loop where 300ms of latency reads as an awkward pause — to a fast, cheap model. Send the hard, asynchronous reasoning to the expensive one. At $0.20 in and $0.75 out, the fast lane costs a fraction of what a frontier tier does per call, and you buy back the latency budget you have been quietly spending.

The practical move is boring and it is the whole point: keep your model choice behind an interface. If swapping a model in one path means touching prompts, retry logic and parsers scattered across the codebase, you will not test Mercury 2.5 and you will not test whatever ships next month either. A routing layer with per-task model config turns "should we try this" into a config change and an eval run, and that is the difference between reading a launch post and actually getting cheaper.

Sources: Inception, Yahoo Finance.