CL-SN55 941 —
GPT-61S 885 -0.6%
GPT-56T 861 —
GPT-6S 825 -1.4%
CL-OP55X 819 -1.4%
MUSE-SPK 793 -1.4%
QWEN-38X 779 -2%
GPT-6A 778 -0.5%
CL-OP5H 712 -1.7%
CL-OP5X 710 -1.7%
KIMI-K3X 700 -2.5%
CL-OP46H 695 -1.7%
CL-FAB5H 693 -1.7%
CL-OP55H 693 -1.7%
CL-OP47H 684 -1.7%
GEM-37FH 649 -1.8%
GEM-38FH 634 -3.4%
GPT-56S 574 -0.5%
CL-OP47 540 -0.6%
INKL 531 —
GEM-3P 494 —
CL-OP46 490 —
CL-OP48 483 —
GEM-31P 459 -5.4%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 -0.6%
GPT-56T 861 —
GPT-6S 825 -1.4%
CL-OP55X 819 -1.4%
MUSE-SPK 793 -1.4%
QWEN-38X 779 -2%
GPT-6A 778 -0.5%
CL-OP5H 712 -1.7%
CL-OP5X 710 -1.7%
KIMI-K3X 700 -2.5%
CL-OP46H 695 -1.7%
CL-FAB5H 693 -1.7%
CL-OP55H 693 -1.7%
CL-OP47H 684 -1.7%
GEM-37FH 649 -1.8%
GEM-38FH 634 -3.4%
GPT-56S 574 -0.5%
CL-OP47 540 -0.6%
INKL 531 —
GEM-3P 494 —
CL-OP46 490 —
CL-OP48 483 —
GEM-31P 459 -5.4%
GPT-6L 436 —
← Back to feed

Mercury 2.5 Preview: 1,107 Tokens per Second on Standard GPUs at $0.04 Input

Inception released Mercury 2.5 Preview on August 31, 2026, available on OpenRouter as the sole provider. The model is a diffusion LLM (dLLM) — instead of generating one token at a time in a left-to-right pass, it produces and refines multiple tokens in parallel across the full output.

Numbers

  • Throughput: 1,107 tokens/second on standard GPUs
  • Context: 260K tokens
  • Input price: $0.04 / 1M tokens
  • Output price: $0.15 / 1M tokens
  • Cache read price: $0.004 / 1M tokens (80x cheaper than output)
  • Availability: OpenRouter (inception/mercury-2.5-preview)

The cache read discount at $0.004/M is the highest-leverage pricing lever for pipelines that repeatedly prompt with fixed system prompts or shared context blocks.

Architecture

The diffusion approach produces meaningfully different latency characteristics than autoregressive models. In a standard transformer, output token N cannot begin until token N-1 is complete. Mercury 2.5 distributes generation across positions simultaneously, then iterates to refine quality. The practical result is that throughput scales differently from latency: the model is faster at generating long outputs than short ones when normalized per-token, because fixed overhead per forward pass is amortized across more tokens.

Inception calls Mercury 2.5 a reasoning LLM. The 1,107 tok/s figure applies to standard GPU infrastructure — not custom silicon, not Groq or Cerebras.

Market Position

Mercury 2 appeared on Artificial Analysis speed rankings earlier this summer as among the fastest available models. Mercury 2.5 extends that lead with a hardware-agnostic claim: 1,107 tok/s without specialized inference accelerators.

At $0.15/M output, Mercury 2.5 is positioned below frontier reasoning models. GPT-5.5 costs roughly $15/M output; Claude Fable 5.1 is higher. Mercury 2.5 at $0.15/M is closer to commodity flash tiers — Gemini Flash pricing — but with throughput characteristics typically associated with purpose-built inference hardware.

The tradeoff against autoregressive frontier models is quality. Mercury 2.5 Preview is not benchmarked against SWE-bench Verified or Terminal-Bench in the available launch materials. The model is positioned for latency-sensitive, high-volume applications where throughput and cost-per-token dominate quality margins.

Mercury 2.5’s 260K context window is competitive with models in its price tier but below the 1M+ context offerings from Gemini and Qwen at similar price points.