Inception AI has released Mercury 3, and the company says its new flagship is faster than two well-known rivals and wins most of the head-to-head tests. The Mercury 3 diffusion model beat OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 on eight of 12 benchmarks, according to Inception’s own numbers, while generating text four to eight times faster. As first reported by SiliconANGLE, the launch is meant to answer the biggest question hanging over diffusion-based language models: can they scale?
Key Takeaways
- Inception AI released the Mercury 3 diffusion model, claiming it outperforms GPT-6 Luna and Claude Haiku 5.5 in speed and benchmarks.
- Mercury 3 generates responses faster by drafting text upfront and refining it in parallel, achieving 1,068 tokens per second.
- The model is designed to address scaling issues; it features a 1 million token context window and is over 40 times larger than its predecessor, Mercury 1.
- Initial tests show mixed results across benchmarks, with Mercury 3 excelling in coding but trailing in some other areas; independent evaluations are still pending.
- Mercury 3 pricing is 20 cents per million input tokens and 70 cents for output, temporarily discounted, but it generally costs more than its competitors.
Table of contents
How the Mercury 3 diffusion model differs from standard LLMs
Most large language models, including GPT-6 Luna and Haiku 5.5, write one token at a time, each depending on the ones before it. Mercury works differently. It drafts the entire response up front, then refines every position in parallel across a few passes.
That design is the source of the speed claims. Inception measured the Mercury 3 diffusion model at 1,068 tokens per second on Nvidia GPUs. In the same comparison, Haiku 5.5 reached about 240 and GPT-6 Luna about 130. Those figures line up with the four-to-eight-times range the company advertises.
The scaling question Inception wants to settle
Co-founder and CEO Stefano Ermon says skeptics have raised the same objection since Mercury 1: whether diffusion can scale. His view is that the limits come from compute and data, not from some deep flaw in the algorithm. Mercury 3 is Inception’s evidence for that argument.
The company hasn’t published a parameter count for the Mercury 3 diffusion model. It says only that the model is more than 40 times the size of Mercury 1. The context window has grown to 1 million tokens, up from 260,000 in Mercury 2.5. Ermon also said training on the next model is already under way.
What the benchmarks show for the Mercury 3 diffusion model
The widest gap came in coding. On SWE-Bench Atlas, Mercury 3 scored 51.1, against 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. That’s a large margin, and software development is a workload where many enterprises are already spending money on AI.
The picture isn’t one-sided, though. On DeepSWE, a second coding benchmark, Mercury 3 trailed Haiku 5.5, 46.3 to 51.6. It also fell behind on one long-context test, and GPT-6 Luna took two other benchmarks by less than a point each. Eight wins out of twelve is a strong showing, but it isn’t a sweep.
The bigger caveat is who ran the tests. These are Inception’s benchmarks, and no outside group has published independent results on the model yet. Buyers should treat the headline numbers as a claim to verify, ideally against their own prompts and data.
Where speed changes the product
Inception’s early customers are building things where lag is visible to users. Kuan Zou, founder and CEO of real-time search startup Octen AI, pointed to smart glasses, where a five-second answer feels broken. In Octen’s tests, median generation time with the Mercury 3 diffusion model was 0.9 seconds.
RevisionDojo, a tutoring app, reports a similar experience. Co-founder and CTO Michael Tong said hints and marked answers now come back in about a second. In both cases, the value is less about raw throughput than about keeping an interactive product feeling instant.
Pricing and access
The Mercury 3 diffusion model lists at 20 cents per million input tokens and 70 cents per million output tokens. Cached input is 90% cheaper. A launch promotion cuts both rates in half for now, which would put the model at 10 cents for input and 35 cents for output.
GPT-6 Luna and Haiku 5.5 each list at 10 cents and 50 cents for typical prompts. So at standard rates, Mercury 3 costs more than both. During the promotion it undercuts them on output and matches them on input. Teams should model costs at the full price, since the discount is temporary.
Access is straightforward. Developers can use Inception’s own platform, or reach the Mercury 3 diffusion model through OpenRouter, Baseten and Vercel. Code already written for OpenAI’s API needs only minor changes, which lowers the cost of running a side-by-side trial.
The business behind the model
Ermon founded Inception in 2024 with fellow computer science professors Aditya Grover and Volodymyr Kuleshov. Last November, Menlo Ventures led a $50 million round, with Microsoft’s M12 venture arm among the backers. A Reuters report said Microsoft was in talks to acquire the startup as of May, and that Inception wanted more than $1 billion. Reuters also said SpaceX had approached the company. No deal with either has been announced.
Read Next
A few related pieces worth your time:
- Next Xbox Price Could Hit $1,800, Report Says
- 8 Lower Middle Market M&A Advisors for Technology Companies in 2026
- The Global HR Stack: How to Architect a Unified Employee Experience Using the Best Employer of Record Platform
What to watch next
Three things will tell us whether this launch matters beyond Inception’s own charts. First, independent evaluations of the Mercury 3 diffusion model on coding and long-context work. Second, whether the speed holds up on production workloads rather than benchmark runs. Third, how the pricing looks once the promotion ends.
- Independent results: none have been published yet.
- Cost at full price: higher than both rivals on typical prompts.
- Next model: training is reportedly already under way.
For technology leaders, the practical move is a small, controlled test on latency-sensitive features such as search, tutoring or assistants. If the Mercury 3 diffusion model delivers its claimed speed without giving up quality on your own tasks, the architecture argument gets a lot harder to dismiss.
Source: SiliconANGLE
