Inception launches Mercury 3, promising over 1,000 tokens per second

US AI startup Inception on Monday launched Mercury 3, a language model built on diffusion rather than the usual word-by-word approach. The company says the model beats OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 across a range of benchmarks while generating more than 1,000 tokens per second. It is available now through Inception's own API, OpenRouter, Baseten and Vercel. All performance figures come from Inception and its customers. No independent measurements have been published.

What Inception launched

Inception published the announcement on its blog on 19 October. Co-founder Stefano Ermon, who is also a professor at Stanford, posted the launch on X at 18:42 Oslo time.

According to Inception, Mercury 3 is more than 40 times larger than the first Mercury model. The company says results across a range of standard benchmarks have tripled since Mercury 1 (mini), while speed has stayed the same. Measured speed is 1,068 tokens per second on Nvidia GPUs, and Inception writes that the model runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5. The context window has grown to one million tokens, up from 260,000 in Mercury 2.5.

List pricing is $0.20 per million input tokens and $0.70 per million output tokens. During the launch period Inception offers a 50 percent discount. Cached input tokens are 90 percent off. The announcement does not say how long the launch discount lasts.

The model is compatible with OpenAI's API. Inception says this means it can be swapped into existing systems without rewrites. For enterprise customers, the company promises dedicated capacity, autoscaling, compliance controls and configurable data retention.

Why speed is the point

Conventional language models write text one piece at a time. Diffusion models work on the whole answer at once and refine it over several passes. Inception has bet that this approach cuts latency.

The company is aiming Mercury 3 at uses where delay is noticeable: browser agents, customer support over chat and voice, real-time search and digital tutoring. When an AI agent has to take many steps in a row, waiting time adds up. Inception argues that a faster model lets an agent finish sooner or try more approaches.

Several early customers are quoted in the announcement. TinyFish AI says Mercury 3 completed half of 50 live web navigation tasks in the company's own test harness, more than GPT-6.1-Sol. Octen AI reports a median generation time of 0.9 seconds in its tests. Customer support company Lorikeet says the model was faster than other models and scored higher both in its evaluations and in production. These are customer claims about their own workloads, not independent tests.

What the announcement does not say

Inception refers to "a range of benchmarks", but the announcement text does not name which benchmarks the comparison with GPT-6 Luna and Claude Haiku 5.5 rests on. The company also does not state the parameter count, where data is processed, or whether an EU data residency option exists. The announcement does not say the model will get open weights.

Our assessment is that Mercury 3 is most interesting where response time costs money or customers, and where tasks are well defined. For demanding analysis or important decisions, it is too early to say whether the model holds up.

What business leaders should do

• Organisations running customer support, voice bots or internal agents where response time is a problem can ask their team to test Mercury 3 against their current model on their own examples. Measure quality, error rate and cost per resolved case, not just speed.
• Clarify data location, retention and the data processing agreement before sending personal or customer data to the service. The announcement does not answer these questions.
• Budget on list price, not the launch discount, when assessing long-term use.

Because the model uses the same API format as OpenAI, the barrier to testing is low, and switching back is just as easy.

Sources and media

• Inception, "Diffusion at Scale: Introducing Mercury 3" (Inception Team), dated 19 October 2026, time of day not stated on the page: Inception's blog
• X post by @StefanoErmon (Stefano Ermon, co-founder of Inception, verified account), 19 October 2026 at 18:42 Oslo time: Stefano Ermon's X post
• X post by @StefanoErmon on pricing and availability, 19 October 2026 at 18:42 Oslo time: Stefano Ermon's X post on pricing and availability
• Sources were checked on 19 October 2026.
• Illustration: AI-generated with Grok Bot / hogby.ai