Inception Launches Mercury 3 AI Model With Speeds Exceeding 1,000 Tokens per Second

Artificial intelligence company Inception has launched Mercury 3, its latest language model designed to deliver high-quality responses at exceptionally high speeds. Announced on October 19, 2026, the model represents a major development in the company’s efforts to demonstrate that diffusion-based language models can compete with conventional AI systems. Inception claims Mercury 3 can match or outperform speed-optimised models such as OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 across several benchmarks while operating approximately four to eight times faster. The company reports generation speeds of up to 1,068 tokens per second on NVIDIA GPUs. These figures come from Inception’s own evaluations and have not been independently verified across all the tasks cited.

Unlike conventional language models that generally generate text one token at a time, Mercury 3 uses a technique known as diffusion-based text generation. This approach begins with an initial representation and progressively refines parts of the output, allowing multiple tokens to be processed in parallel. Inception says this architecture enables the model to maintain high generation speeds even as its capabilities and size increase. According to the company, Mercury 3 is more than 40 times larger than the original Mercury 1 model, while benchmark performance has improved approximately threefold compared with Mercury 1 Mini. The latest release also expands the model’s context window to one million tokens, allowing it to process large amounts of text, extensive documents and lengthy conversation histories within a single context.

One of the model’s main targets is the growing market for AI agents that need to make repeated decisions while interacting with websites, applications and external tools. Inception says early partners are already testing Mercury 3 for browser navigation, real-time search, interactive education, customer support and data engineering. In one company-reported evaluation involving 50 live Mind2Web website-navigation tasks, Mercury 3 successfully completed 50% of the tasks using the same token volume and number of agent steps as the comparison model. Inception said the result exceeded GPT-6.1 Sol in that particular test. Faster responses can be especially valuable for autonomous agents because delays accumulate when a system must repeatedly analyse a page, select an action and evaluate the result. However, performance in a limited benchmark does not establish that Mercury 3 is superior for every agent workflow.

Mercury 3 also introduces a pricing strategy aimed at developers and businesses seeking to reduce the cost of running AI-powered applications. Inception lists standard pricing at $0.20 per million input tokens and $0.70 per million output tokens, with a launch promotion reducing those rates to $0.10 and $0.35 respectively. The company has also offered 100 million free tokens to help developers evaluate the model. Mercury 3 is available through Inception’s API platform and several distribution services, including OpenRouter, Baseten and Vercel. Its API is designed to be compatible with the OpenAI API format, potentially allowing developers to integrate the model into existing applications without completely rewriting their systems. This compatibility could make it easier for smaller companies and independent developers to experiment with faster AI models.

The launch adds another dimension to the competition between AI developers, which has largely focused on improving reasoning, coding performance, multimodal capabilities and model reliability. Mercury 3 challenges the assumption that faster AI systems must sacrifice substantial intelligence, although its claimed advantages will need to be evaluated independently across a wider range of real-world tasks. Early testing by external developers has already suggested that its speed may be particularly useful for shorter coding tasks and agent workflows, while longer, multi-step coding sessions can still expose limitations. This distinction is important because rapid text generation does not automatically translate into better reasoning or more reliable results. Businesses will need to consider accuracy, consistency, security and total operating costs when choosing a model for production use.

Inception’s Mercury 3 launch highlights the growing diversity of architectures being explored across the AI industry. If diffusion-based language models continue improving while retaining their speed advantages, they could become important components of real-time assistants, voice applications, coding tools and autonomous agents. Their ability to produce responses quickly may also allow systems to test more possible actions within the same time and computing budget. Nevertheless, the long-term significance of Mercury 3 will depend on independent evaluations, sustained performance under heavy demand and its ability to handle complex tasks reliably. For now, the release gives developers another option for building fast AI applications and adds pressure on established AI companies to improve the speed and affordability of their own models.