# Mercury 3 launches on Venice, bringing parallel token generation to private AI > Source: > Published: 2026-10-22 15:25:12+00:00 # Mercury 3 launches on Venice, bringing parallel token generation to private AI Inception Labs' diffusion-based model claims speeds above 1,000 tokens per second and is now accessible through the privacy-focused Venice platform Most chatbots write the way a nervous student fills in an exam: one word at a time, never looking back. Inception Labs is betting that approach is a bottleneck. The company’s new model, **Mercury 3**, is now live on Venice.ai. It uses a diffusion architecture that generates and refines every token in a response at once, pushing output speeds past 1,000 tokens per second. ## What Mercury 3 actually does Inception Labs launched Mercury 3 on October 19, 2026. It is the fourth model in the company’s Mercury series. The headline figure is speed. Inception says the model hits 1,068 tokens per second when running on [NVIDIA](https://cryptobriefing.com/markets/nvidia/) GPUs. Mercury 3 drafts an entire response simultaneously, then sharpens all of it in parallel passes. Speed is not the only spec. Mercury 3 ships with a 1-million-token context window, meaning it can hold a very large amount of text in working memory during a single task. It also supports tool calling, structured outputs and tunable reasoning. Put simply, it can trigger external software, return answers in tidy machine-readable formats, and dial its thinking effort up or down. ## Benchmarks, pricing and the fine print Inception claims Mercury 3 outperforms or matches [OpenAI](https://cryptobriefing.com/markets/openai/)’s GPT-6 Luna and [Anthropic](https://cryptobriefing.com/markets/anthropic/)’s Claude Haiku 5.5 on 8 to 9 out of 12 benchmarks. According to Inception, the model shows particular strength in instruction following and agent-based coding. It reportedly does this while running at 4 to 8 times the speed of those rivals. Those benchmarks are internal, though. Independent validation is still pending. ### AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. Inception also says Mercury 3 is more than 40 times larger in parameters than Mercury 1. Despite that growth, the company claims it keeps similar speed and efficiency. There is one early real-world data point. An early partnering agent reportedly completed 50% of tasks on Mind2Web, a benchmark built around navigating websites to finish multi-step jobs. On pricing, list rates are set at $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount cuts that to $0.10 and $0.35 respectively. Access is API-only, with compatibility for OpenAI’s systems. Developers can call it using the model ID “mercury-3”. OpenAI compatibility means teams already built around OpenAI’s tooling can, in theory, swap in Mercury 3 without rewriting their entire stack. ## Why Venice is part of the story Mercury 3 is being offered through privacy-focused platforms, including Venice.ai, under a pay-per-token model. The integration landed shortly after launch. ## How we got here The Mercury line has been iterating quickly. Prior versions arrived across 2025 and 2026, with each release pushing the diffusion concept further. ## What this means Inception is targeting a specific slice of the market: low-latency work. That includes AI agents, real-time search, and voice and chat interactions. For developers, the pitch is straightforward. Comparable quality to small frontier models, several times faster, at launch prices of $0.10 and $0.35 per million tokens. The risk is in the verification. Self-reported benchmarks have a habit of looking better in the press release than in production, and the 8-to-9-out-of-12 claim has not yet been independently confirmed. The discount is also temporary. Teams that build around the $0.10 and $0.35 rates should budget for the list prices of $0.20 and $0.70 once the promotion ends. Things to watch next: independent benchmark results, how Mercury 3 handles long tasks that actually use its 1-million-token window, and whether the agent performance seen on Mind2Web holds up beyond an early partner. **Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).