Mercury 3 Is a Diffusion Language Model That Writes 1,068 Tokens a Second

For two years, the AI industry has treated speed and intelligence as a tradeoff you cannot beat: you pick a fast model or a smart one. On October 19, Inception decided that rule is optional. The Redwood City startup launched Mercury 3, a diffusion language model that generates text at 1,068 tokens per second while claiming benchmark scores on par with or better than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5.

That number is worth pausing on. In Inception's own testing, GPT-6 Luna generated about 130 tokens per second and Claude Haiku 5.5 around 240. Mercury 3 clocks in at 1,068, on ordinary NVIDIA GPUs, while doing up to eight times the work in the same window. If the claim holds, the fast-but-dumb tier of AI models just became a lot less dumb.

Why the Mercury 3 Diffusion Language Model Matters

Most language models write like a typewriter: one token at a time, each word conditioned on everything before it. No matter how fast your hardware gets, a long answer takes a long time. That is the autoregressive approach, and it runs everything from ChatGPT to Claude.

Diffusion models take the opposite path. Instead of committing to each token in order, they draft the entire response at once, then refine every position in parallel over several passes. It is the same idea behind image diffusion models that resolve noise into a picture, applied to text. The parallelism is what buys the speed. The open question, since the very first diffusion language models, has been whether that speed can survive at scale.

Inception's answer is emphatic. Mercury 3 is more than 40 times larger than Mercury 1, launched in June 2025, yet runs at comparable speeds. "We're showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers," said co-founder Stefano Ermon. The company has already started training the next model. That is either a genuine architectural breakthrough or the bravest bluff in AI this year, and the next six months of independent testing will decide which.

What 1,068 Tokens a Second Actually Changes

Token speed alone sounds like a benchmark brag. It is not. The real story is agents. When an AI agent completes a task, it calls the model over and over: navigate, read, think, act, think again. Small delays compound at every step. A model that responds in one second instead of eight can try more approaches before answering, or finish the task before the user loses patience.

Early-access partners are already pointing at this. Education platform RevisionDojo says Mercury 3 returns hints, marked answers, and definitions in about a second, calling it a drop-in swap for its existing stack. Search company Octen AI reported a median generation time of 0.9 seconds and end-to-end response time of 2.2 seconds. Browser-use agents tested by TinyFish AI on 50 live web navigation tasks completed half of them, a result Inception says beat GPT-6.1-Sol with the same token budget and step count.

These are vendor-published partner claims, not independent results. Treat them as directions to test, not conclusions. But the underlying logic is sound: in the race to make agents fast and cheap, raw generation speed is the lever nobody else is pulling.

The Benchmark Claims, and What's Missing

Across its own benchmark suite, Inception says Mercury 3 beat GPT-6 Luna and Claude Haiku 5.5 in 8 of 12 tests at four to eight times their speed. The widest lead came in coding: 51.1 on SWE-Bench Atlas against 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. Mercury 3 trailed Haiku 5.5 on DeepSWE, a second coding benchmark, 46.3 to 51.6, and GPT-6 Luna won two tests by less than a point each.

Here is the catch that matters for developers: no outside group has published independent results on Mercury 3 yet. And there is a subtle wrinkle in how the model presents itself. Mercury 3 has four reasoning-effort settings (instant, low, medium, high) and defaults to low, where its predecessors defaulted to medium. A model answering at low effort will look faster and score lower than the same model at high effort, and Inception's comparison does not spell out which setting produced each number. If you benchmark it yourself, pin the effort level explicitly or your numbers will not be comparable to anyone else's.

This is the part of the speed-tier model boom where you stay skeptical and run your own numbers.

Mercury 3 Pricing and Where to Try It

Mercury 3 is available today through the Inception API, Baseten, OpenRouter, and Vercel. It is OpenAI API compatible, so teams can point their existing code at a new base URL without rewriting anything. Enterprise deployments add dedicated capacity, autoscaling, compliance controls, and configurable data retention.

List pricing is $0.20 per million input tokens and $0.70 per million output tokens. During the launch window, both are half price: $0.10 and $0.35, with cached input tokens at a 90% discount. The company has not said when the introductory discount ends, so treat the cheap prices as temporary, not structural. The 1-million-token context window, up from 260,000 in Mercury 2.5, gives developers room to feed in long documents without chunking, though long inputs still cost money at every token.

One deadline to note: teams on Mercury 2 or 2.5 have until November 15, 2026 to keep access; after that, migration is mandatory. Enterprise customers remain covered under their agreements.

Who Should Pay Attention

Mercury 3 is not for everyone. If your workloads are about deep reasoning with no latency pressure, a bigger autoregressive model remains the safer choice. But if you are building anything where the model gets called in a loop (support agents, tutoring, live search, voice), Mercury 3 is worth a serious look this week. The list price is competitive, the launch discount makes it nearly free to evaluate, and the OpenAI compatibility means the experiment costs an afternoon, not a rewrite.

The deeper question is whether diffusion scaling laws are real. If they are, every lab with a fast-model strategy just found a new playbook, and the AI capacity crunch gets a very interesting plot twist. If they are not, Mercury 3 is still a fast, cheap model with a million-token window, which is not the worst thing to be. Either way, it is the most interesting model release of the month, and independent testers will have the final word.