Inception introduced Mercury 3 on October 19, 2026. The diffusion language model has a stated one-million-token context and a company-reported throughput of 1,068 tokens per second on NVIDIA GPUs. Announced pricing is $0.20 per million input and $0.70 per million output tokens, with a 50% introductory discount. API access is available. These are vendor results, not SIExpo tests.

Speed has several different meanings

Readers should separate fast text generation from fast completion of useful work. Our example is an assistant that compares spare-part availability, checks delivery conditions and prepares an offer. If one lookup returns the wrong product, a rapidly written description does not make the offer usable. Conversely, a correctly selected part can be valuable even when the accompanying response is shorter.

That is our evaluation framework, rather than a report of Mercury operating in such a store. A comparison would start with identical requests and an explicit completion criterion: for example, the right item with a verifiable availability source. It should record both time to the first visible response and time to a result that actually meets that criterion. Those measurements answer different questions about the user experience.

Check integration against the actual workflow

The official quick start shows the mercury-3 model identifier, reasoning_effort and max_completion_tokens settings, and access through the Inception SDK or compatible client libraries. This provides a basis for planning a comparison. An example request does not prove that every existing agent integration will preserve its behavior without changes.

In our proposed workflow, an evaluation would examine missing stock information, an interrupted external lookup and a request needing clarification. Sometimes the useful next step is a question to the customer rather than another confident offer. The success criteria should allow that outcome; otherwise, a comparison may reward verbosity instead of correct handling of uncertainty.

Cost also needs a common denominator. We recommend comparing spending per successfully completed request under the same constraints, accounting for repeated model calls and separate external operations. The introductory discount should be shown separately from the announced standard rate. That prevents a temporary promotion from becoming an unstated assumption in a longer-term assessment.

For a meaningful test, the organization should preserve failed cases as well as successful ones. A faster run that requires manual repair may have a different value from an equally fast run that completes correctly. Recording the repair work would reveal that distinction without assuming which model will perform better in advance.

Mercury 3 adds a concrete option for applications sensitive to delay. A deployment decision still requires local measurements of quality, complete workflow time and final cost. The advertised generation rate is relevant evidence about one part of that assessment; it is not a substitute for the assessment itself.

Editorial verification: 2026-10-20T06:10:23.481994+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.