Inception Labs' Mercury 2.5 claims 1,107 tok/s diffusion LLM speed at $0.75 per million output tokens

A speed comparison published on dev.to names Inception Labs' Mercury 2.5 the fastest LLM reachable through an API as of September 2026. The diffusion-based model posts a vendor-reported 1,107 tokens per second on widely available NVIDIA GPUs, while OpenRouter's live telemetry records 440 tok/s at the median with about 1.17 seconds of latency — enough to lead every rival in the roundup on both figures. At $0.20 per million input tokens and $0.75 per million output tokens, it also offers the cheapest output pricing of the group.

How the four models stack up

The dev.to piece measures Mercury 2.5 against Gemini 3.5 Flash-Lite, GPT-5.6 Luna and Claude Haiku 4.5. Using Artificial Analysis figures cited there, Flash-Lite runs at 382 tok/s and costs $0.30/$2.50 per million; Luna reaches up to 129 tok/s at $0.20/$1.20 with the largest context window of the four; Haiku 4.5 manages roughly 82 tok/s in non-reasoning mode at $1.00/$5.00. On output price, Haiku costs about 6.7 times Mercury and Flash-Lite about 3.3 times.

The verdict splits by workload. Latency-bound production traffic points to Mercury; cheap, long-context general chat to Luna; Google-native stacks to Flash-Lite; and Haiku 4.5 earns its place through Anthropic's instruction-following rather than throughput.

Why diffusion gets there

Conventional autoregressive models such as Luna and Haiku emit one token at a time, so total response time grows with output length. Mercury is a diffusion language model: it refines a block of tokens in parallel across denoising steps, separating speed from sequence length — the origin of Inception's 1,107 tok/s claim.

Inception shipped the first commercial diffusion LLM in February 2025 and followed with Mercury 2 in February 2026. The company, founded in 2024 by Stanford's Stefano Ermon, UCLA's Aditya Grover and Cornell's Volodymyr Kuleshov, raised a $50M seed led by Menlo Ventures with Microsoft's M12 and NVIDIA's NVentures participating, plus angel backing from Andrew Ng and Andrej Karpathy. Inception also claims a 40% intelligence gain over Mercury 2 and says a larger model is already in training.

What deployments show

Two vendor-published customer figures stand out. Augment Code reports compaction latency falling from 150 seconds to 27 seconds — an 82% reduction — with cost down 90%. Voice-agent firm OpenCall reports median response latency near 170 ms and P99 dropping from minutes to about one second. The dev.to author flags both as vendor numbers rather than independent benchmarks, so treat them as directional.

OpenRouter's traffic pattern fits the story: the endpoint's top consumers are agentic clients, with Hermes Agent at 3.86 billion tokens and Claude Code at 2.3 billion — workload shapes where a tool makes many short internal calls and the user's wait is the total of all of them.

Caveats worth budgeting for

First, the headline speed is Inception's own hardware figure; the 440 tok/s median observed on OpenRouter is closer to what most callers will see, so benchmark on your own prompts. Second, the 80% launch discount that priced Mercury at $0.04/$0.15 expired on 8 September 2026 — plan against list prices. Third, claimed quality parity with the cheap tier rests on Inception's own eval suite, and practitioners note its speed comparisons reference older competitor releases. Fourth, the 260K context window, while plenty for agent loops, is the smallest here and matters for whole-repository passes.

On integration, the API is OpenAI-compatible and supports tool calling, parallel tool calls, structured outputs and tunable reasoning. It is also served through Baseten and OpenRouter, which lists maximum output at 65,536 tokens, and Inception offers 100 million free API tokens to start.

Why it matters

For latency-bound applications — voice agents, search pipelines, coding subagents — throughput and tail latency, not peak intelligence, decide whether a product feels usable, and Mercury 2.5 currently leads measured speed while charging the lowest output price among the compared models. Diffusion's parallel decoding also changes the economics of agent loops, where per-call latency compounds across many short requests. The honest read is narrower than the headline: observed throughput sits well below the vendor figure, capability claims are effectively self-graded, and context is limited. The dev.to piece's suggested approach — replaying a few hundred real production prompts across all four models and aggregating at P50 and P99 rather than the mean — is the sensible way to test whether the speed advantage survives contact with actual workloads.