Inception
Inception's Mercury 3 diffusion model tops 1,000 tokens a second
Inception released Mercury 3, a diffusion language model with a 1M-token context window, at a launch price of $0.10 per million input tokens.
Inception released Mercury 3, a language model it says matches the quality of fast frontier models while generating text four to eight times faster.
Inception calls Mercury 3 “the most capable diffusion-based LLM released to date” in a blog post dated Oct. 19, 2026. The company contrasts diffusion models with autoregressive models, which write one token at a time.
Speed and Quality
Inception reports 1,068 tokens per second on NVIDIA GPUs. The company says Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running four to eight times faster. The post describes quality as “on par or better than” speed-optimized frontier models. The comparison is Inception’s own, and the post names no independent evaluation.
The context window is 1M tokens, up from 260K for Mercury 2.5.
Price
Mercury 3 lists at $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount that brings it to $0.10 and $0.35. Cached tokens cost 90% less than the input price. Those are the prices Inception posted on Oct. 19. OpenRouter’s model list showed the discounted $0.10 and $0.35 rates and a 1M-token context for inception/mercury-3 on Oct. 20.
Scale
Inception says Mercury 3 is more than 40 times larger than Mercury 1 and about three times more intelligent on standard benchmarks than Mercury 1 (mini), “just as fast.” The company says the result shows text diffusion scales.
Customer Tests
TinyFish AI cofounder Shuhao Zhang said Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in a production web-agent harness, “significantly more than GPT-6.1-Sol.” He said the run used the same token volume and number of agent steps as a frontier model.
Octen AI founder Kuan Zou said Mercury 3 had a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds in tests of Octen’s Sense API. He called it the best balance of speed and quality among the fast models Octen tested. Search itself took about 80 milliseconds.
RevisionDojo CTO Michael Tong said hints, marked answers and definitions come back in about a second after a drop-in swap. Early-access partners also use Mercury 3 for voice and chat support, real-time tutoring and data engineering.
Availability
Mercury 3 is available on the Inception API Platform and through OpenRouter, Baseten and Vercel. It is compatible with the OpenAI API.
Enterprise deployments offer dedicated capacity, autoscaling, compliance controls and configurable data retention. Inception offers 100 million free tokens to try the API.
Mercury 3 is the flagship of a real-time lineup that includes Mercury Voice, introduced Sept. 29, 2026, and Mercury Decide, a decision model for high-volume classification.
Our View
For developers building chat or voice assistants, the speed figure sets how long a customer waits on the model. By our math, a 500-token reply at 1,068 tokens per second takes about half a second to generate. We think that puts the model’s share of the wait at well under a second.
Developers with an OpenAI-style integration can try Mercury 3 on a copy of the assistant, using Inception’s 100 million free tokens. They can then read its replies to their own customer conversations beside the current model’s.
Every comparison in the post comes from Inception or its partners. The speed helps customers only if those replies are as good as the ones the current model gives.