A Diffusion Model Now Writes 1,100 Tokens a Second, Cheaply

Mercury 2.5 Preview appeared on OpenRouter late on August 31, 2026 and on Inception’s own site the following day. It is the latest in the company’s line of diffusion language models, which generate text by producing and refining many tokens in parallel rather than predicting them one at a time. Inception’s claim is specific: 1,107 tokens a second on standard GPUs, a ten-point jump in intelligence over Mercury 2, and quality comparable to the cost-optimised frontier tier it names as GPT-5.6 Luna at low effort, Gemini 3.5 Flash-Lite and Claude Haiku 4.5. The list price is $0.20 per million input tokens and $0.75 per million output, and an 80% discount runs until September 8 at 07:00 UTC.

Every one of those numbers is the vendor’s, and this is a preview. The point of the post is not to repeat them but to say what they would mean if they hold, which workloads they change, and how to check them in the week the discount gives you.

A conventional language model is autoregressive: it predicts the next token, appends it, and predicts again, so generating a thousand tokens means a thousand sequential passes and the hardware spends most of its time waiting on memory rather than computing. A diffusion language model starts from a rough draft of the whole output, or a large block of it, and refines every position at once over a fixed number of steps. If the number of refinement steps is much smaller than the number of tokens, the model finishes far sooner, and it uses the GPU the way GPUs like to be used, on wide parallel work. That is the whole trick, and it is why Inception’s first Mercury models, covered in our February post on Mercury 2, already claimed four-figure token rates.

The historical cost was quality. Refining in parallel makes it harder to keep long-range dependencies straight, and early diffusion models trailed autoregressive ones of similar size on reasoning. Inception’s framing for 2.5 is that the gap to the cost-optimised frontier tier has closed; Google’s open-weight DiffusionGemma, covered in our June post, made a similar argument for the research community. Neither claim has an independent benchmark attached at the time of writing.

“1,107 tokens/sec on standard GPUs” is a real measurement of something, and the four questions to ask of it are the ones we set out in how to read an AI speedup claim: which hardware, what batch size, what output length, and whether it is peak or sustained. Inception answers none of them on the page. The most useful independent signal is OpenRouter’s live provider table, which at the time of writing showed the Inception endpoint delivering 477 tokens a second at 3.56 seconds of latency. That is a different measurement, taken on mixed real traffic with unknown prompt shapes, and it is still roughly four to five times faster than typical autoregressive small models on the same table. Both figures can be honest. The gap between them is the gap between a lab benchmark and a shared endpoint, and your own number will land somewhere between, depending on output length.

The quality claim is harder to check, because Inception states a “10+ point jump” without naming the index, and a comparison set without scores. Until an independent index publishes a row for Mercury 2.5, the right posture is the one we apply to every launch table: treat “comparable to Haiku 4.5” as the vendor’s hypothesis and run it against your own evaluation set.

Throughput only changes a decision where time, not tokens, is the binding constraint. Three workload shapes qualify. The first is the agent loop with many short generations: a planner that calls tools twenty times per task spends most of its wall-clock time waiting for the next short response, and a model that returns it in a fraction of a second changes how long the task takes more than any change in per-token price. The second is anything with a human waiting on the other end, where sub-second responses are the difference between a feature people use and one they abandon. The third is bulk rewriting, classification and extraction over large corpora, where the job is measured in hours and a five-fold throughput gain is a five-fold shorter job.

Where it does not matter is the long, hard reasoning task, where quality decides and the user will wait. Mercury 2.5’s own comparison set, the cheapest tier from each frontier lab, tells you where Inception thinks it belongs. That tier is exactly where our cost-per-task benchmark of headless coding agents found the widest spread between models that finish and models that merely respond, so speed alone is not a reason to route there.