Mercury 3 promises shorter waits, with a launch discount and a migration date

An assistant that answers quickly but needs another three rounds of correction has not necessarily saved time. Inception’s October 19 Mercury 3 launch is aimed at that waiting problem, with a one-million-token context window and a vendor-reported 1,068 tokens per second on NVIDIA GPUs. Treat that throughput figure as a measurement supplied by Inception, not a promise for your complete workflow. Dated announcement.

The company describes Mercury 3 as a diffusion language model. It is offered through Inception’s platform and several serving partners. Its API quick start uses the mercury-3 identifier with an OpenAI-compatible chat-completions interface. That familiar request shape is a place to begin integration; it does not establish that every existing tool or parser will behave identically. Official quick start.

Keep the discount separate from the budget

The current model card shows launch prices per million tokens of $0.10 input, $0.01 cached input and $0.35 output, beside list rates of $0.20, $0.02 and $0.70. One million uncached input tokens plus one million output tokens would therefore be $0.45 at launch rates or $0.90 at list rates. This is illustrative arithmetic, not a workload we ran. The opened launch post does not supply a discount end date, and a serving partner can have different terms.

For a proposed creator assistant, keep a small set of real questions and source documents fixed. Record when the first useful answer appears, when the whole answer finishes and whether a person can accept it without repairing factual mistakes. Include any search or tool delay. Fast model output cannot remove a slow external lookup, and raw throughput alone cannot decide which assistant feels responsive. We have not performed this comparison.

Existing users have a date to check

Inception’s model page says Mercury 2 and 2.5 are closed to new users. Customers with usage before October 19 retain access until November 15, 2026, while enterprise agreements remain supported under their terms. That is a reason to schedule a migration check, not assume every account has the same cutoff.

Start with one workflow and keep a fallback while comparing the result. Our guide to models, apps and agents helps identify which part of the stack you are changing.