OpenRouter's P50 puts Mercury 2.5 at 440 tok/s against Inception's reported 1,107

A diffusion language model refines a block of tokens in parallel across denoising steps, so throughput stops tracking sequence length, and Inception's 1,107 tok/s comes out of that ^[8]. Luna and Haiku emit one token at a time, so their latency grows with the length of the output ^[9].

Block-parallel decoding only speeds up the generation part of a call. OpenRouter records 1.17 s of latency alongside its 440 tok/s median for the endpoint ^[3]. A 300-token reply at 440 tok/s is about 0.68 s of generation ^[6]. On short replies the fixed 1.17 s is the larger half of the wall clock, and the architecture does nothing to it.

On long outputs the gap is wide. A thousand output tokens take about 2.3 s at 440 tok/s and about 12.2 s at Haiku's measured 82 tok/s, a difference near 10 s per call ^[3]. Luna's 129 tok/s puts the same thousand tokens at 7.8 s ^[4].

The four speed figures in the dev.to comparison, last verified 10 September 2026, do not come from one rig ^[17]. Mercury's 440 tok/s is OpenRouter production telemetry ^[3]. Flash-Lite's 382 and Luna's 129 are Artificial Analysis numbers ^[4][5]. Haiku's roughly 82 is described as measured non-reasoning ^[6]. Inception's 1,107 is its own figure on widely available NVIDIA GPUs ^[1]. Against Flash-Lite, the vendor number is 2.9x ^[7] and the OpenRouter median is 1.15x ^[2]. A buyer who wants the ratio that applies to their own service has to run both endpoints on one harness with their own output-length distribution.

Output price is the one axis where Mercury's lead does not depend on whose harness you trust: $0.75 per million output against Flash-Lite's $2.50 and Haiku's $5.00 ^[2][4][6], which the comparison puts at 3.3x and about 6.7x ^[16]. The 80 percent launch discount that ran to 8 September held input and output at $0.04 and $0.15 ^[7], so a pilot costed inside that window was costed at a price that no longer exists. List price is five times both figures ^[5].

On capability the record is thinner. Inception reports a 40 percent intelligence gain over Mercury 2 and calls Mercury 2.5 the most capable diffusion LLM on the market ^[10]. The comparison's four speed figures have third-party sources; its capability claim for Mercury 2.5 is the vendor's own. Its recommendation for general cheap chat is Luna, on the grounds that a larger context window than Mercury's 260K and the maturity of the OpenAI ecosystem matter more than tokens per second when a human is reading the output ^[11][2].

The traffic fits the narrow case. OpenRouter's top consumers of the endpoint are agentic clients, Hermes Agent at 3.86B tokens and Claude Code at 2.3B, a shape where a coding tool fires many short internal calls and the user waits on the chain ^[12]. Two vendor figures point the same way: Augment Code says compaction latency fell from 150 s to 27 s with cost down 90 percent ^[13], and OpenCall reports median response latency around 170 ms with P99 down from minutes to about a second ^[14]. The comparison labels both as vendor numbers, not independent benchmarks ^[13].

The port itself is cheap. Mercury exposes OpenAI-compatible tool calling, parallel tool calls and structured outputs, which the comparison calls a drop-in for existing agent code ^[15]. The test worth running before the swap is your own median output length against the 1.17 s in OpenRouter's telemetry.