Inception Labs releases Mercury 2.5, a diffusion LLM that runs at 1,107 tokens per second on standard NVIDIA GPUs

A coding agent that waits four seconds between each tool call is a different experience from one that waits half a second. Inception Labs published a number on that gap on Wednesday.

The company published Mercury 2.5 on 24 September, its latest diffusion language model, and priced it at $0.20 per million input tokens and $0.75 per million output tokens on its own API. A launch discount cuts that to $0.04 and $0.15 per million for a limited window. The independent benchmark site Artificial Analysis measured the API at 770.4 output tokens per second, third out of the 174 models it tracks, against a median of 111.6 for the group.

Inception says the model runs at 1,107 tokens per second on standard NVIDIA GPUs and that median response latency for a voice agent built on it sits close to 170 milliseconds. Those are the vendor's own numbers, so treat them as the company's claim rather than a verified figure. The Artificial Analysis measurement, made against the live API, is closer to what a customer will see.

What is different about the model

Mercury 2.5 is a diffusion language model rather than the autoregressive kind used by most named LLMs. In an autoregressive model, the output arrives one token at a time, each token conditioned on the ones before it. A diffusion model refines a full block of tokens in parallel, so the wall-clock time to finish a response scales differently. Inception calls it "the largest diffusion language model ever trained" and says quality is comparable to cost-optimized frontier models: GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. Those are vendor-selected comparisons; nobody outside the company has yet benchmarked Mercury 2.5 head to head with them on quality.

The Augment Code case study, with numbers

Inception's launch post carries one named-customer example. Augment Code, a coding-agent company, moved its context compaction step to Mercury 2.5 and reports that the latency for that step fell from about 150 seconds to 27 seconds, and the cost fell by 90 percent. Tool-search summaries, Augment says, now return in under a second. These figures are Augment's, filtered through Inception's marketing, so the honest reading is that one customer measured a large gain on one workflow, and other teams would need to test the same swap themselves.

The model is available through Inception's own API, and also through Baseten and OpenRouter, so a team can route calls to it without leaving an existing gateway. A hosted playground at mercury.inceptionlabs.ai lets a developer test prompts before writing any integration.

For a team building an agent that makes many small model calls, the price and speed together matter more than either alone. A workflow that runs 500 tool-response summarisations per user session on a model at $3 per million output tokens looks different if the same call costs 25 cents per million and returns four times sooner. Whether Mercury 2.5 gives comparable answers on the specific task is the question every team will have to answer for itself, because the vendor's own quality comparisons carry weight only until an independent test measures the same workload.

Source

• Introducing Mercury 2.5 at Inception Labs
• Mercury 2.5 benchmark page at Artificial Analysis