Inception Says Mercury 3 Diffusion Model Tops GPT-6 Luna at 4–8x the Speed

Inception on 19 October 2026 launched Mercury 3, the latest in its line of diffusion large language models, with a 1-million-token context window and output speeds the company puts above 1,000 tokens per second. Inception claims that across a range of benchmarks Mercury 3 outperforms OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 while running four to eight times faster. For developers building agents and real-time apps, where every model call adds latency, that would mean a more capable model without the usual wait.

The comparison is Inception’s own. The release names no independent evaluation, so the benchmark wins should be read as company claims until outside testers check them. What is confirmed is the launch itself, the price list and availability on four platforms.

Why diffusion speed matters

Most LLMs are autoregressive: they write one token at a time. Diffusion models generate many tokens in parallel, which is why Inception has pitched them as a faster route to inference. The open question, the company says, was whether that speed would survive as the models grew more capable.

Inception’s answer is scale. It says Mercury 3 is more than 40 times the size of Mercury 1, which launched in June 2025, yet runs at comparable speeds on widely available NVIDIA GPUs. Mercury 3 is the fourth release since then: Mercury 2 added reasoning, Mercury 2.5 added tunable reasoning and a 260,000-token context, and Mercury 3 expands that to one million tokens.

Co-founder Stefano Ermon said the model is “over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second”, arguing that the limits on scaling diffusion are “compute and data, not fundamental algorithmic barriers”. He added that Inception has already started training its next model.

Price, platforms and early users

The model is OpenAI API compatible, so teams can switch without rewriting code, Inception says. Enterprise deployments offer dedicated capacity, autoscaling, compliance controls and configurable data retention. Vercel’s changelog also lists Mercury 3 on its AI Gateway.

Inception says early-access partners are testing it in browser-use agents, interactive education, real-time search and customer support. Michael Tong, co-founder and CTO of General Learning (RevisionDojo), said in the release that hints, marked answers and definitions “now come back in about a second”, calling it “a drop-in swap for our existing stack”. That quote was supplied through Inception’s announcement.

What the release does not show

The release gives no benchmark names, scores or methodology for the GPT-6 Luna and Haiku 5.5 comparison, only the claim that Mercury 3 wins “across a range of benchmarks”. Speed figures depend on hardware and workload, and the release does not say which GPUs or batch sizes produced the 1,000-tokens-per-second number. Mercury 3 is also not described as open-weight; it is offered through APIs.

Our read

Mercury 3 is the strongest case yet that diffusion LLMs can grow without giving up their speed edge, and the low launch price makes it cheap to test. But a headline win over GPT-6 Luna and Haiku 5.5 with no published scores is a marketing claim, not a result. The next useful signal is third-party benchmarks and real latency numbers from OpenRouter and Vercel users — and whether Inception’s next model, already in training, keeps the curve going.

Sources

1. Inception: “Inception Launches Mercury 3, Setting a New Pareto Frontier in Speed and Intelligence” (Business Wire, 19 October 2026).
2. Vercel changelog: “Mercury 3 now available on AI Gateway” (19 October 2026).

Related stories

• DeepSeek Releases DeepSeek-V4.1-Flash-1017 Open-Weight Update
• MiniMax Rolls Out M3.1 Flash, a Faster, Cheaper Agentic Model