Most language models write the way a person types: one word after another, in order. Mercury 3, which Inception announced on October 19, does something different. It drafts a whole answer at once and then refines it, in parallel. According to Inception, it now does that at 1,068 tokens per second on NVIDIA GPUs, fast enough that the company is pitching it against OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5, the models labs sell on speed.

Mercury 3 is the fourth release in Inception’s Mercury line since Mercury 1 appeared in June 2025, and it is by far the biggest. The company says the model is more than 40 times larger than Mercury 1, yet runs at roughly the same speed. The context window grows to 1 million tokens, up from 260,000 in Mercury 2.5. For a technology that began as an academic curiosity (early diffusion models could do little more than autocomplete code), those are striking numbers, if they hold up.

“Since we launched Mercury 1, the question we’ve heard most is whether diffusion can scale,” said Stefano Ermon, Inception’s co-founder. “Mercury 3 is over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second. We’re showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers. We’ve already started training our next model, and we expect to continue pushing intelligence forward without sacrificing speed.”

Abstract illustration of a diffusion model refining many tokens in parallel
Illustration created for GPTOnline.

What the benchmark table actually says

Inception’s own comparison table, reproduced in MarkTechPost’s launch coverage, has Mercury 3 ahead of GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks. The standout is SWE-Bench Atlas, a software engineering test, where Mercury 3 scores 51.08, about 18 points clear of both rivals. The table is less flattering elsewhere: on DeepSWE v1, Mercury 3 scores 46.31 against Haiku 5.5’s 51.62. In other words, this is a speed story with competitive quality, not a clean sweep, and every figure in that table comes from Inception itself. None of it has been independently verified yet.

The early field reports are, so far, about responsiveness rather than raw intelligence. Michael Tong, co-founder and CTO of the tutoring platform RevisionDojo, said in the launch announcement: “Mercury 3 allowed us to give every student instant feedback without giving up reasoning. Hints, marked answers and definitions now come back in about a second, with quality as good as or better than the models we ran before. When a student is stuck on a question, that speed keeps them working instead of waiting, and it was a drop-in swap for our existing stack.”

Abstract illustration of fast, repeated model calls inside an AI agent workflow
Illustration created for GPTOnline.

Price, platforms and the fine print

Mercury 3 is live now on the Inception API, Baseten, OpenRouter and Vercel, and it speaks the OpenAI API format, so existing code mostly carries over. OpenRouter’s listing shows launch pricing of $0.10 per million input tokens and $0.35 per million output tokens, half the standard rates of $0.20 and $0.70, with cached input at $0.01 per million tokens, a 1-million-token context window and support for tool calling. Inception says early workloads include browser agents, real-time search, customer support and interactive education: exactly the jobs where a model is called again and again, and where a slow answer compounds into a sluggish product.

There are two caveats worth keeping in view. First, the speed and benchmark claims are vendor numbers, measured on Inception’s chosen tests; the honest test is what independent evaluators find on workloads Inception did not pick. Second, the launch quietly retires the older models for new users: according to Unite.AI’s coverage of Inception’s models page, Mercury 2 and Mercury 2.5 are no longer available to new sign-ups, with existing customers keeping access until November 15, 2026.

None of that diminishes the underlying bet. The autoregressive approach has had a decade-long head start and the full weight of the industry’s engineering behind it. Inception is arguing that generation speed is not a fixed tax on intelligence, and with Mercury 3 it has put a specific, checkable number on that claim: 1,068 tokens per second.

Sources