Mercury 2.5: 1,107 tokens per second from a diffusion LM, and a 5x price jump the day the promo ended

Inception's Mercury 2.5 Preview is a diffusion language model, and the vendor-run number attached to it is 1,107 tokens per second. That is a different way of decoding, not a small model on a big GPU. The other number attached to it landed today. At 07:00 UTC on September 8 the launch discount of $0.04 per million input tokens and $0.15 per million output expired, and the list price is $0.20 / $0.75. Anyone who routed a fast-cheap tier onto Mercury during the first week is paying 5x more this morning for the same tokens. Both numbers deserve a closer look, because they pull in opposite directions.

Why diffusion decodes faster

An autoregressive model produces one token per forward pass. Every token depends on the previous one, so the sequence is inherently serial, and tokens per second is bounded by how fast you can run one pass. Speculative decoding shaves this by guessing ahead, but it still verifies serially.

A diffusion language model starts from a block of noise or masked tokens and refines the whole block over a number of steps. Each step is a parallel operation over every position, so the number of forward passes depends on the step count, not the token count. If you can converge a block of a few hundred tokens in a handful of steps, the effective throughput is enormous, and that is where a four-digit tokens-per-second figure comes from. The general trade-off, before anyone's marketing enters the picture: fewer refinement steps means more speed and more risk that the block did not converge, and parallel generation turns the left-to-right dependencies that autoregression handles for free into something the model has to learn.

Diffusion buys throughput by trading serial certainty for parallel refinement; the price of that trade is set by your quality eval, not by the tokens-per-second chart.

What to measure, separately

• Prefill and decode as two numbers. Diffusion changes decode; it does not change the cost of reading your prompt. A headline tok/s on a short prompt and a long completion tells you nothing about a RAG workload that is 90% prompt.
• Quality-adjusted speed. Tokens you have to regenerate are not throughput. I laid out the method in quality-adjusted token speed; run it on your own prompts before believing any tok/s figure, vendor or otherwise.
• Latency to first usable output, not first token. A diffusion model may emit its whole block at once, which is great for batch and irrelevant for streaming UIs.
• Behaviour on structured output. Strict JSON and code have exactly the kind of long-range constraints that parallel refinement finds hardest. Test them explicitly.
• Cost per completed task at list price, not at the promo. Then compare to the live price table.

When throughput is worth it after the promo

At $0.04 / $0.15, Mercury was cheap enough that speed was a free bonus. At $0.20 / $0.75 it sits in the same band as the aggressive small models, and this week DeepSeek moved V4 Flash into public beta with a vision variant at no surcharge, so the cost-sensitive tier is crowded. Throughput now has to justify itself on its own.

• It does for batch pipelines with a wall-clock deadline: nightly classification, bulk rewriting, synthetic data generation.
• It does for interactive tools where the whole response is needed before the next step, such as an agent generating a complete file to compile.
• It does not for streaming chat, where a human reads at a few tokens per second and anything above that is invisible.
• It does not if your quality eval shows regeneration rates that eat the speed.

If you use a router, the practical lesson is about promo pricing rather than diffusion. A launch discount is a routing signal with an expiry date. Put the expiry in the routing config as a rule, not in someone's memory.

The honest limit

The 1,107 tok/s is Inception's own measurement on Inception's hardware with an undisclosed prompt and completion mix, and the model is a preview. I have not seen an independent benchmark, and I have deliberately kept the architectural description general because I do not know which of the many diffusion-LM design choices Mercury makes. What I can say from the price sheet alone is that a 5x jump on day eight is a reason to re-run every cost model that assumed day one.