The Pulse
Inception Launches Mercury 3 at 1,068 Tokens per Second
Inception Labs launched Mercury 3 on October 19, 2026, claiming generation speeds of 1,068 tokens per second on NVIDIA GPUs. The diffusion-based model offers a one-million-token context window, discounted launch pricing and access through I
AI.info Team · Written with AI · fact-checked ·
Mercury 3 allowed us to give every student instant feedback without giving up reasoning.
Michael Tong, Co-founder/CTO, General Learning aka RevisionDojo
Mercury 3’s headline speed: 1,068 tokens per second
Inception Labs launched Mercury 3 on October 19, saying its new diffusion-based language model generates 1,068 tokens per second on NVIDIA GPUs. The company says it runs four to eight times faster than the speed-focused models it compares against, while matching or exceeding their quality across a range of benchmarks. Those comparisons name GPT-6 Luna and Claude Haiku 5.5.
The speed figure is Inception’s own reported result, not an independent test presented in the announcement. The post gives benchmark scores in a table, where Mercury 3 leads on eight of the twelve benchmarks listed, but not a detailed account of the setup behind its broad model comparisons. It also quotes one browser-agent evaluation by early-access partner TinyFish AI: Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in TinyFish’s production web-agent harness, using the same token volume and number of agent steps as a frontier model. TinyFish co-founder Shuhao Zhang says that completion rate was significantly higher than GPT-6.1-Sol’s, but the post does not give the comparator’s score.
A larger model with a one-million-token context
Inception frames Mercury 3 as evidence that diffusion models can scale beyond smaller, specialized systems. The company says the model has more than 40 times Mercury 1’s total parameters and scores three times higher across a range of standard benchmarks than Mercury 1 Mini. A chart in the post measures that rise as “Model Quality”, an average of eleven benchmarks including GDPval, GPQA Diamond and BrowseComp+.
Mercury 3’s context window holds one million tokens, up from 260,000 in Mercury 2.5. Diffusion language models use a different generation approach from the familiar left-to-right process: Mercury’s earlier product material describes proposing and refining multiple tokens in parallel. Inception pitches that architecture for tasks where delays accumulate across repeated model calls, such as browser agents and conversational support.
Launch pricing starts at $0.10 per million input tokens
Inception lists standard prices of $0.20 per million input tokens and $0.70 per million output tokens. At launch, it discounts both rates by half, to $0.10 and $0.35 respectively; cached input tokens receive a 90% discount against the input price. The announcement does not give an end date for the launch offer.
The model is available through Inception’s API Platform, OpenRouter, Baseten and Vercel. Inception says it supports the OpenAI API format and offers 100 million free tokens through its API. Enterprise deployments can include dedicated capacity, autoscaling, compliance controls and configurable data retention.
Early users put the speed claim to work
Inception points to early customers in tutoring, web navigation, search and customer support. RevisionDojo’s Michael Tong says hints, marked answers and definitions return in about a second, and describes Mercury 3 as a drop-in replacement for the company’s existing model stack. Those are customer-reported outcomes; the launch post does not include an independent evaluation of the deployments.
For Octen AI’s search tests, founder and CEO Kuan Zou reports a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds. The figures show why throughput alone is not the whole performance story: Octen also measured the time for search and model generation together. Mercury 3’s practical value will depend on how its claimed speed and benchmark results hold up on customers’ own workloads, especially where answer quality and full-response latency matter as much as tokens per second.