Inception Labs Launches Mercury 3, Its Bet That Diffusion Can Rival GPT-6
Inception Labs released Mercury 3 on October 19, positioning it as the most capable diffusion-based large language model shipped to date and the clearest test yet of whether an architecture built around parallel generation can compete with token-by-token autoregressive models like GPT-6 and Claude.
Most commercial LLMs generate text one token at a time, predicting each word from everything before it. Mercury's diffusion approach instead starts with a rough draft of the full output and iteratively refines it in parallel passes, a method Inception has argued trades some of autoregressive modeling's precision for dramatically higher throughput. With Mercury 3, the company is now claiming that diffusion models obey the same scaling laws that have driven autoregressive gains for years — a claim with real stakes, since it implies diffusion isn't just a niche speed optimization but a path that keeps improving with scale.
The numbers, and who's checking them
Inception says Mercury 3 has more than 40 times the total parameters of the original Mercury and that measured performance on its internal benchmark suite roughly tripled over that span, though it has not disclosed an absolute parameter count. The model ships with a 1-million-token context window, up from 260,000 in Mercury 2.5, and a claimed generation speed of 1,068 tokens per second on Nvidia GPUs — 4 to 8 times faster than what Inception calls comparable models.
None of that has been independently confirmed. Artificial Analysis, a third-party benchmarking firm, has previously measured Mercury 2 at roughly 680 tokens per second and Mercury 2.5 at roughly 706 — well below Inception's own figures for those models — but as of October 20 had not yet published a Mercury 3 measurement at all. Inception's claims that Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 "across a range of benchmarks" likewise come with no published scores attached.
Partner results carry the same caveat
The strongest evidence offered so far comes from Inception's own commercial partners rather than neutral evaluators. TinyFish AI, which builds web-browsing agents, reported that Mercury 3 completed 50% of 50 live Mind2Web navigation tasks — notably more than GPT-6.1-Sol at the same token and step budget, the partner said, though it published no score for the rival model either. Octen AI, which uses Mercury for smart-glasses assistance, reported a median 0.9-second generation time and 2.2-second end-to-end response in its own tests. Tutoring platform RevisionDojo and customer-support vendor Lorikeet both praised the model's latency in statements supplied for Inception's launch materials.
That pattern — benchmark claims validated only by companies with a commercial stake in Mercury's success — is the piece analysts and outside commentators have focused on since launch, rather than disputing the architecture itself.
Pricing undercuts the speed story's economics: Inception is charging a launch rate of $0.10 per million input tokens and $0.35 per million output tokens, roughly half its own list price and, by its own framing, about a fifth of Claude Haiku 5.5's output rate. The company is also sunsetting Mercury 2 and 2.5 for new signups, giving existing customers until November 15 to migrate.
Mercury 3's real test won't be Inception's launch post — it will be whether Artificial Analysis or another outside lab publishes numbers that hold up, and whether other diffusion-LLM efforts follow Inception's scaling bet rather than treating the architecture as a speed-only sideshow to mainstream autoregressive models.