Inception Labs introduced Mercury 3 on Monday, describing it as the most capable diffusion-based large language model released to date. According to the company's announcement, Mercury 3 outperforms speed-optimized frontier models including GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4 to 8 times faster, at more than 1,000 tokens per second.
The launch is the strongest answer yet to the question that has hung over commercial diffusion language models since the category emerged: can text diffusion scale to compete with autoregressive architectures on complex workloads, or does it remain a research curiosity optimized only for latency? Inception Labs, which was founded by former Stanford and UC Berkeley researchers including CEO Stefano Ermon, says the answer is now settled. The company reports that Mercury 3 is more than 40 times larger than Mercury 1 and that measured intelligence across standard benchmarks roughly tripled between Mercury 1 (mini) and Mercury 3, evidence, it argues, that the scaling laws which drove progress in autoregressive models hold for diffusion too. For more context on this story, see our ongoing more AI stories.
Speed, Context and Pricing
The headline figures are aggressive. Inception Labs says Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs, maintaining the signature advantage of the diffusion approach over token-by-token autoregressive decoding. The context window has grown to 1 million tokens, up from 260,000 in Mercury 2.5, which the company released in September with 1,107 tokens per second of sustained output.
Mercury 3 is priced at $0.20 per million input tokens and $0.70 per million output tokens. At launch, the model is 50 percent off, bringing the effective rate to $0.10 per million input tokens and $0.35 per million output tokens, with cached tokens priced 90 percent below the input rate. That undercuts the typical pricing of the frontier speed-optimized models it benchmarks against, though independent verification of the quality claims has not yet been published.
Benchmark and Agent Results
Beyond the general benchmark comparisons against GPT-6 Luna and Claude Haiku 5.5, Inception Labs published results from its production web-agent harness on 50 live Mind2Web navigation tasks. Mercury 3 completed 50 percent of the tasks, which the company says is significantly more than GPT-6.1-Sol managed, while using the same token volume and the same number of agent steps as the frontier model at a fraction of the per-token cost.
The company positions latency-sensitive agent workloads as the core commercial case for the model. Because browser-use agents, voice assistants, and interactive applications wait on the model at every step, small delays compound quickly, and a diffusion model's high throughput translates directly into faster task completion and more attempts within a fixed time budget.
Early Customers Report Production Use
Inception Labs says early-access partners are already running Mercury 3 in production for voice and chat support, real-time tutoring, data engineering, and web navigation.
Shuhao Zhang, cofounder of TinyFish AI, said agents built on Mercury 3 finish sooner and have time to try more approaches before answering, whether they are navigating websites, researching across sources, or analyzing logs and tracing incidents.
Michael Tong, co-founder and CTO of General Learning, which operates RevisionDojo, said Mercury 3 allowed the company to give every student instant feedback without giving up reasoning quality, returning hints, marked answers, and definitions in about a second, and described the model as a drop-in swap for its existing stack.
Kuan Zou, founder of search startup Octen, reported that in the company's Sense API tests its own search took about 80 milliseconds while model generation was the bottleneck, and that Mercury 3 delivered the best balance of speed and quality among the fast models tested, with a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds.
Why Diffusion Scaling Matters
Autoregressive transformers, which generate text one token at a time in sequence, have carried essentially every frontier model from GPT through Gemini and Claude. Diffusion models instead refine an entire response in parallel through iterative denoising, a property that makes generation fast and controllable but which, until recently, lagged badly on complex reasoning tasks.
Mercury 3's central claim is that this quality gap has closed at scale. If the company's numbers hold up under independent evaluation, the practical consequence is a credible second architecture family for commercial inference, with implications for the economics of high-volume agent workloads, where aggregate token spend is dominated by speed-tier models.
The claims remain vendor-reported for now. Inception Labs has not pointed to third-party evaluations of Mercury 3, and its benchmark comparisons come from its own harnesses. The company's earlier models, however, have been publicly served since 2025, giving buyers a track record to check against.
Availability
Mercury 3 is available immediately through Inception Labs' API at the launch-discounted rate, with the company stating that cached tokens are billed at 10 percent of the input price. No open-weights release was announced; like its predecessors, Mercury 3 is a commercial, API-only model.
For continuing coverage of new model releases, diffusion architectures, and the widening speed tier of frontier AI, follow AI Buzz Wire.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
