Mercury 3 launches a faster, cheaper diffusion model for agents
Throughput above one thousand tokens per second could shorten agent loops, but temporary pricing and unverified results do not yet justify a full migration.
What shipped
Inception introduced Mercury 3 on October 19 as a diffusion language model with a one-million-token context window. The company reports 1,068 tokens per second, and access is live through its API, OpenRouter and Vercel AI Gateway.
Why it matters
Latency compounds when an agent calls a model repeatedly. If Mercury 3 preserves sufficient quality at this speed, it could shorten search, voice and browser workflows without forcing teams to redesign their entire stack.
What we would do
We would not replace a primary model on vendor claims alone. Virtual Arc would shadow-route production-shaped tasks, budget against list rather than launch pricing, and require a lower cost per successful task before expanding traffic.
Mercury 3 is interesting enough for an immediate controlled pilot, but not proven enough to become our default production route. If Inception’s claim of 1,068 tokens per second survives real workloads, the important gain will not be a faster-looking chat window; it will be shorter completion times for agents that call a model dozens of times per task. Virtual Arc would not build a business case around the promotional price of $0.10 per million input tokens and $0.35 per million output tokens, because no end date has been published. We would budget at the $0.20 and $0.70 list rates, then shadow-route Mercury 3 across voice, search, browser automation and coding subagents. We would measure time to completed task, success rate, retries and total execution cost rather than raw token throughput. Until independent testing and our own evaluations confirm quality, reliability and sustained performance under load, we would keep the incumbent model as a fallback. This is a credible specialist accelerator, not yet a reason for a wholesale migration.