← Back to feed

Inception released Mercury 3 on October 16, and the claim on the tin is blunt: the model is more than 40 times larger than the original Mercury while generating at the same per-second throughput. That is an unusual thing to be able to say about a scale-up.

The throughput figure is 232 tokens per second at median latency, with a p50 first-token time of 0.88 seconds. That speed comes from architecture, not hardware tricks. Mercury models are diffusion-based — they produce and refine tokens in parallel rather than one at a time. The larger the model, the more expensive each generation step is, but the parallelism scales differently than autoregressive inference. What Inception is claiming is that their scaling did not erode the throughput advantage that made Mercury interesting to begin with.

What It Costs

Mercury 3 launches at $0.10 per million input tokens and $0.35 per million output tokens, a 50 percent launch discount from the full-price rate of $0.20 and $0.70. Cache reads are $0.01 per million at the discounted rate. Inception is hosting it directly — OpenRouter routes every request to a single provider — and the model carries 99.97 percent uptime over the trailing 30-day window.

For comparison, Mercury 2.5 has been one of the fastest models on OpenRouter, measured in raw tokens per second. Mercury 3 gives up some of that raw speed in exchange for a large capability step. At $0.10 per million tokens, it undercuts most frontier reasoning models by a substantial margin while remaining in the speed tier that makes sub-second agent turns practical.

Context and Capabilities

The context window is one million tokens, with a maximum of 131,072 completion tokens. Mercury 3 supports tool calling through the standard tools and tool_choice parameters, structured outputs via a JSON schema in response_format, and streams normally.

Inception describes the primary use cases as search agents, voice pipelines, and coding subagents — workloads where latency compounds across multiple sequential calls. A multi-step research agent that calls an LLM ten times per query is much more sensitive to per-call latency than a single-shot summarization task. At 0.88 seconds median first token, Mercury 3 is designed for that compounding environment.

The model is not yet available through EU in-region routing, which matters for enterprise deployments with data residency requirements.

Where Diffusion LLMs Stand

Inception has built a line of diffusion models — Mercury, Mercury 2, Mercury 2.5, Mercury Decide — each iterating on the same architectural premise. The autoregressive models that dominate the frontier (GPT, Claude, Gemini, Llama) generate tokens sequentially: each token depends on the last, which imposes a fundamental throughput ceiling regardless of hardware. Diffusion sidesteps that constraint by treating generation as an iterative refinement over the full output, enabling parallelism.

The catch has always been quality. Early diffusion language models were fast but uncompetitive on standard benchmarks. Inception has not released benchmark scores for Mercury 3, so direct quality comparisons are not yet possible. What the release does establish is that the scaling trajectory is intact: Inception has grown the model by 40x without sacrificing the speed characteristic that distinguishes it.

If Mercury 3 proves competitive on agent tasks — where speed often matters more than peak reasoning depth — it positions Inception as a viable infrastructure layer for production AI pipelines that need both throughput and capability at a cost below frontier list prices.