Inception is positioning Mercury 3 as a language model for applications where waiting at each step can slow down the whole experience. The company says its diffusion approach helps it generate text faster than conventional models, while its own benchmark results show a mix of wins and losses against GPT-6 Luna and Claude Haiku 5.5.
How Mercury 3 generates answers
Many language models produce text one word at a time, with each next word depending on what came before. Inception describes Mercury 3 as working differently: it starts with a rough version of the whole answer, then refines it over several passes. That lets it generate multiple words at once.
Diffusion models are better known for generating images. The same broad idea—begin with an incomplete version and refine it repeatedly—is being applied here to language. Google’s technical report on DiffusionGemma explains the basic idea, according to the article.
Inception calls Mercury 3 the most capable diffusion language model released so far. It says the model generates 1,068 tokens per second on Nvidia GPUs, compared with about 240 for Claude Haiku 5.5 and about 130 for GPT-6 Luna in the company’s comparison. Those speed figures have not been independently tested.
Benchmark results vary by task
In Inception’s benchmark table, Mercury 3 leads in eight of 12 tests when all three models run at medium settings. Its strongest reported advantages include instruction following and autonomous coding. On IFBench, it scores 76.84 points, 15 more than GPT-6 Luna. On SWE-Bench Atlas, it scores 51.08 points, while its rivals score 33.33 and 32.26.
The table does not show Mercury 3 winning every category. Claude Haiku 5.5 performs better on the longer coding test DeepSWE and at finding information in long texts. GPT-6 Luna is ahead in two tests from Artificial Analysis. Inception’s shorter overview describes Mercury 3’s quality as on par or better, a more measured claim than a clean sweep.
Inception also says Mercury 3 has more than 40 times as many parameters as Mercury 1 while running just as fast. Performance has tripled since the small Mercury 1 mini, according to the company. The exact parameter count has not been disclosed.
Designed for apps where delays accumulate
Inception is targeting AI agents, search, customer support and tutoring. An agent may need to call a model repeatedly as it works through a task, so even modest waits can add up across multiple steps. Faster generation could make those interactions feel more responsive, if the speed holds in real deployments.
Early-access partners cited in the announcement emphasize that responsiveness. Octen AI, which builds search features for smart glasses, says the search itself took only about 80 milliseconds and the language model was the bottleneck. With Mercury 3, it reports a median response after 0.9 seconds and a total process time of 2.2 seconds.
Education company General Learning says it can now deliver hints and corrections to students in about one second. Customer service provider Lorikeet reports faster response times along with better results. TinyFish AI says Mercury 3 completed 25 of 50 navigation tasks on real websites, beating GPT-6.1-Sol; co-founder Shuhao Zhang did not give a score for the OpenAI model.
Pricing, access and the limits of the claims
Mercury 3 can process up to 1 million tokens at once. Its regular listed pricing is $0.20 per million input tokens and $0.70 per million output tokens, with Inception cutting those prices in half at launch. The company has not said when the discount ends.
After the discount, output would cost more than GPT-6 Luna and Claude Haiku 5.5 for shorter inputs, where both charge $0.50. For very long prompts, both rivals charge considerably more, according to the source article. The large context window may therefore matter when comparing costs, alongside the model’s speed and task performance.
The model is available on Inception’s platform, OpenRouter, Baseten and Vercel. It uses an OpenAI-compatible API, which Inception says should make switching easier for developers. Still, the central speed and quality comparisons remain company-reported; independent testing will be needed to establish how Mercury 3 performs across real workloads.