Inception Mercury 3 Diffusion Model Claims 8X Speed Advantage Over GPT-6 Luna and Claude Haiku 5.5

Inception Releases Mercury 3 Diffusion Model With Speed-First Architecture

Inception, the Redwood City AI lab, released Mercury 3 on Monday, marking its fourth model since launching Mercury 1 in June 20251

1

. The newest version represents what the company calls a diffusion-based large language model, a design that fundamentally differs from the autoregressive models powering most chatbots today. While traditional systems write one token at a time, each chosen sequentially after the last, the diffusion model works more like image generators. It starts with rough, noisy output and refines many tokens simultaneously in parallel until the text resolves1

1

. Inception says this parallel approach explains why its model runs faster than competitors.

Mercury 3 Outperforms GPT-6 Luna Across Benchmarks at 1000 Tokens Per Second

On the company's own benchmarks, Inception Mercury 3 beats OpenAI's GPT-6 Luna, Anthropic's Claude Haiku 5.5, and Zhipu's GLM-5.3 Flash across a range of tests while running four to eight times faster1

1

. Inception says it generates more than 1000 tokens per second on Nvidia chips, though these figures are self-reported and no outside group has published independent testing yet2

2

. The lab says Mercury 3 is more than 40 times the size of its first model and about three times as capable across standard benchmarks, at the same AI model speed1

1

. "Since we launched Mercury 1, the question we've heard most is whether diffusion can scale. Mercury 3 is over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second," said Stefano Ermon, co-founder of Inception2

2

.

One Million Token Context Window Enables Low-Latency Interactions

The context window, the amount of text the model can handle at once, has grown to one million tokens, up from 260,000 in the previous version1

1

. This one million token context window nearly quadruples the capacity introduced in Mercury 2.5, allowing the model to process substantially larger amounts of information in a single request2

2

. The expanded window matters most for applications requiring low-latency interactions, including interactive tutoring, live translation, writing assistants, and automated software agents. In agent-based systems, models often perform multiple sequential tasks such as searching websites, analyzing information, or examining software logs. Delays at each stage can increase the time required to complete an operation, making speed a critical factor2

2

.

Launch Pricing Undercuts Rivals With Cost-Effective AI Strategy

Mercury 3 costs $0.20 per million input tokens and $0.70 per million output tokens. For launch, the company is halving both rates to $0.10 and $0.35, though Inception has not said how long that promotion will last1

1

. The launch price undercuts GPT-6 Luna and Claude Haiku 5.5 on output tokens, extending a wider run of price cuts across the industry. Speed, not price, remains the primary pitch. "It's comparable to the Luna and Haiku models, but it's way faster," Inception chief executive Stefano Ermon told The Deep View. "It's getting 1,000 tokens per second. It's almost instant in the way you get answers from these models"1

1

. A model that answers faster is easier to run at scale, which matters most for agents and coding tools that call a model many times in a row.

Recursive Self-Improvement Powers Model Development

Inception also described how it built the model using an innovative technique. Ermon told The Deep View that the company used swarms of Mercury 3 agents to improve the model itself, a form of recursive self-improvement1

1

. He said the technique added about 16% to the model's speed, on top of the gains his team made by hand. Inception would not disclose how many parameters the model has in total. Early-access customers are already testing the model in production workflows. Educational technology company RevisionDojo reported that Mercury 3 reduced response times for student feedback, including hints, definitions, and marked answers, to approximately one second2

2

.

Availability Across Multiple Platforms With OpenAI Compatibility

The model is available now through Inception's own API, plus OpenRouter, Baseten, and Vercel's AI Gateway1

1

2

2

. It uses the same API format as OpenAI's, allowing developers to integrate it into existing applications without extensive code changes. The compatibility means developers can switch existing code with little modification. Inception's backers include Menlo Ventures, Microsoft's M12, Nvidia's venture arm, and investors Andrew Ng, Andrej Karpathy, and Eric Schmidt1

1

. Diffusion remains a minority bet in the field. Google DeepMind released a diffusion text model, DiffusionGemma, in June, signaling that larger labs are testing the same idea, though most of the field remains focused on autoregressive models1

1

. Independent testing will show whether the speed claims hold up outside Inception's own benchmarks.