Inception says Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs, pitching it for browser agents, interactive tutoring, search and customer support. Shorter waits could add up across repeated model calls, provided the model completes those tasks reliably.
The company’s October 19, 2026 announcement describes Mercury 3 as generally available through its own API, OpenRouter, Baseten and Vercel. It pairs a diffusion architecture with a one-million-token context window, OpenAI API compatibility and launch pricing of $0.10 per million input tokens and $0.35 per million output tokens.
Availability and listed pricing have supporting provider documentation. The headline speed, quality comparisons and customer results remain vendor-reported evidence, giving builders reasons to evaluate Mercury 3 without yet proving its advertised advantage across production workloads.
Agent Latency Is the Real Target
An agent repeatedly calls a model to decide what to do, invokes a tool, reads the result and makes another decision. Generation delays can accumulate across that sequence, even when each response seems reasonably quick.
Mercury 3’s architectural bet is that diffusion can reduce the sequential generation bottleneck associated with autoregressive decoding. Inception claims the model runs four to eight times faster than the speed-optimized models it compares against.
How much that helps depends on how much of an agent’s total runtime comes from model generation. Faster output could materially improve a workflow dominated by long reasoning or generated responses. It will have less influence when the agent spends most of its time waiting for websites, database queries or external APIs.
Throughput also differs from responsiveness. Tokens per second describes output generation speed; it does not, by itself, establish how quickly the first useful answer arrives. For voice interfaces and short tool calls, a fast start may matter more than sustained output speed.
A million-token prompt adds another variable: rapid generation does not establish rapid processing of a full context window.
The Quality Claims Need More Than a Speed Chart
Inception says Mercury 3 matches or exceeds models including GPT-6 Luna and Claude Haiku 5.5 across its benchmark comparisons. It also says Mercury 3 has more than 40 times the total parameters of Mercury 1.
Both are company claims. The launch material does not disclose Mercury 3’s total parameter count, and the supplied evidence does not include an independent reproduction of its benchmark or throughput results. The relative size increase provides development context without revealing deployment requirements or establishing capability on its own.
