Inception Unveils Mercury 3, Claiming Diffusion Model Outpaces Rival AI Systems
Inception AI has introduced Mercury 3, a diffusion language model generating over 1,000 tokens per second while competing with GPT-6 Luna and Claude Haiku 5.5.
Diffusion-based language model startup Inception AI Inc. has introduced its latest flagship model, Mercury 3. According to internal evaluation data shared by the company, Mercury 3 outperformed OpenAI Group PBC’s GPT-6 Luna and Anthropic PBC’s Claude Haiku 5.5 across eight of 12 benchmark evaluations, delivering output at four to eight times the generation speed of its competitors.
Rather than relying on traditional autoregressive generation, Mercury 3 employs a diffusion-based framework. While systems like GPT-6 Luna and Haiku 5.5 generate responses sequentially by predicting one token after another, Mercury models generate an initial draft of the full output simultaneously before refining all token positions in parallel across multiple passes. Running on Nvidia Corp. graphics processing units, Inception recorded Mercury 3 at an output rate of 1,068 tokens per second. In the same comparative setup, Haiku 5.5 achieved approximately 240 tokens per second, while GPT-6 Luna reached roughly 130 tokens per second.
Scaling Diffusion for Text Generation
Inception co-founder and Chief Executive Stefano Ermon addressed persistent questions surrounding diffusion architectures in natural language processing. Ermon noted that the primary inquiry facing the startup since the debut of Mercury 1 “is whether diffusion can scale.” In his assessment, current boundaries are governed by “compute and data, not fundamental algorithmic barriers.”
While Inception has not published an exact parameter count for Mercury 3, the company confirmed that the model is more than 40 times larger than Mercury 1. The context window has expanded to 1 million tokens, marking a substantial increase over the 260,000 tokens supported by Mercury 2.5. Training for the next iteration following Mercury 3 is already in progress, according to Ermon.
On developer evaluations, Mercury 3 demonstrated its largest margin over GPT-6 Luna in software engineering tasks. On the SWE-Bench Atlas benchmark, Mercury 3 posted a score of 51.1, compared to 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. However, Mercury 3 did not lead across every test. It lagged behind Haiku 5.5 on the DeepSWE benchmark (46.3 versus 51.6) and on a long-context evaluation. GPT-6 Luna also edged out Mercury 3 on two additional tests by margins under one point. Independent third-party benchmarks have not yet been published.
Low-Latency Implementations and Benchmarks
Early enterprise deployments of Mercury 3 are focused primarily on latency-sensitive environments. Octen AI, a real-time search provider, has integrated the model into wearable devices. Octen founder and Chief Executive Kuan Zou pointed out that for smart glasses, “an answer that takes five seconds feels broken.” During testing with Mercury 3, Octen observed a median generation time of 0.9 seconds. Similarly, tutoring platform RevisionDojo integrated the architecture, with co-founder and Chief Technology Officer Michael Tong stating that student hints and evaluated answers are returned in roughly one second.
Founded in 2024 by computer science professors Stefano Ermon, Aditya Grover, and Volodymyr Kuleshov, Inception previously secured a $50 million investment round led by Menlo Ventures, with backing from Microsoft Corp.’s M12 venture fund. Reuters previously reported that Microsoft held discussions regarding an acquisition earlier this year, with Inception seeking a valuation above $1 billion, and that Space Exploration Technologies Corp. also explored discussions. Neither firm has completed an acquisition.
What it means for developers
Mercury 3 introduces a distinct structural alternative for development teams prioritizing throughput and interactive latency. The model is priced at $0.20 per million input tokens and $0.70 per million output tokens, alongside a 90% discount on cached inputs. A promotional launch window currently reduces both rates by 50%. By comparison, standard published pricing for both GPT-6 Luna and Claude Haiku 5.5 sits at $0.10 per million input tokens and $0.50 per million output tokens for standard requests.
Developers can deploy Mercury 3 directly through Inception’s platform or via third-party infrastructure providers including OpenRouter Inc., Baseten Labs Inc., and Vercel Inc. Applications currently built around OpenAI’s API specifications require only minor adjustments to connect to Mercury 3 endpoints.
For teams evaluating different model architectures across latency, cost, and task performance, platforms like Apixo provide pay-per-token API access to frontier models including Claude, GPT, Gemini, Grok, and DeepSeek via a single key at https://apixoai.online. As diffusion language models mature alongside standard autoregressive options, developers will have more architectural choices to balance raw execution speed against standard inference pricing.
Source: Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 at up to 8x the speed — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
