Mercury 2.5 hits 770 tokens per second in independent Artificial Analysis testing

Mercury 2.5 posts 770 tokens per second in independent testing

Mercury 2.5, the newest release in Inception Labs' Mercury family of language models, has been measured generating output at roughly 770 tokens per second by Artificial Analysis, the independent benchmarking service. The figure comes from the Artificial Analysis model page for Mercury 2.5, which drew wide attention after reaching the Hacker News front page. According to Artificial Analysis, the number reflects the performance of the model's first-party API — or the median across providers where no first-party API exists — counted as tokens per second delivered while the model is actively generating.

How the benchmark works

Artificial Analysis defines output speed as the rate at which tokens arrive once generation has started, after the first chunk of a streamed response. The service pairs that with several related measurements on the same page: latency, expressed as seconds to the first answer token and including any hidden thinking time for reasoning models; an end-to-end response time computed for a 500-token answer; and decode time per task on the Artificial Analysis Intelligence Index. That index, currently at version 4.3.2, aggregates ten evaluations, among them Terminal-Bench 4.0, SciCode and Humanity's Last Exam, alongside agentic knowledge-work and coding suites the service runs itself. The page also tracks weighted cost per Intelligence Index task, token usage and context window, so the headline speed number sits inside a broader picture of capability, price and latency rather than standing alone.

Where the speed comes from

Mercury is built on a diffusion architecture. Instead of generating text strictly one token at a time, the way conventional autoregressive transformers do, diffusion language models refine entire sections of a response across a fixed number of parallel steps. That design is what makes decode rates in the hundreds of tokens per second technically plausible: much of the output is produced simultaneously rather than sequentially. Hosted frontier models generally deliver output at a small fraction of Mercury 2.5's measured pace, putting the result well outside the normal range for production APIs. At that rate, text arrives far faster than a person can read, which changes what interactive applications can feel like.

What fast decode changes

The consequences fall into two buckets: economics and latency. On the economics side, generation time is GPU occupancy. A model that finishes a response several times faster frees serving capacity sooner, which raises the number of requests a given deployment can handle and puts downward pressure on cost per request. Artificial Analysis frames this explicitly through its cost-per-task metric, which weights input, cached, reasoning and output token prices against the work actually completed.

On the latency side, output speed matters most where responses are long or where calls chain together: voice agents that need natural turn-taking, coding assistants streaming suggestions, and agentic pipelines in which each step's wait compounds. Reasoning models are a special case, since they spend large numbers of tokens on internal thinking before answering — faster decode directly shrinks the wall-clock penalty that thinking incurs.

There are caveats. Time to first token is not the same as output speed, and for reasoning models it can dominate the wait a user actually perceives. Single-stream measurements may also diverge from throughput under heavy concurrent load, and speed is only useful in proportion to the quality of what is being generated — the reason Artificial Analysis publishes speed next to its intelligence and cost indexes rather than as an isolated figure.

Why it matters

Inference speed is becoming a competitive axis in its own right, alongside model quality and price. A production model sustaining around 770 output tokens per second, as Artificial Analysis measured for Mercury 2.5, suggests that diffusion-based serving can move from research demos into commercially relevant territory. If those rates hold at scale and at reasonable cost, real-time voice interfaces, live coding tools and long multi-step agent runs all become cheaper and more responsive — and rivals that still decode one token at a time will face a measurable disadvantage on the stopwatch, not just the leaderboard.