Inception announced Mercury 3 on October 19, 2026, reporting generation throughput of 1,068 tokens per second on NVIDIA GPUs. That is a vendor-reported result, not a guaranteed response time for an application. The model is available through Inception and distribution partners, including Vercel. Source: Inception announcement.
For businesses evaluating AI agents, the decision is narrower than whether a new model looks fast. Does it reduce the time needed to complete an acceptable task under your serving conditions?
What Mercury 3 changes
Inception describes Mercury's diffusion approach as generating tokens in parallel. Its platform provides an OpenAI-compatible API. This makes the architecture relevant to teams exploring responsive applications, but compatibility should not be treated as proof that existing prompts and integrations need no testing. Source: Inception platform documentation.
The model documentation lists a one-million-token context window, tool calling and structured outputs. Standard rates are $0.20 per million input tokens and $0.70 per million output tokens. The published launch promotion halves those rates to $0.10 and $0.35. Vercel also lists the discounted rates. Budget against the regular price rather than assuming a launch offer is permanent. Source: model specifications; Vercel listing.
Throughput is not the whole latency measurement
Vercel distinguishes live generation throughput from time to first token in its model dashboard. Those are separate measurements, and neither by itself represents the entire time required to finish a business task. Source: Vercel metric definitions.
Our recommendation is to measure three things independently: when the first useful output arrives, when the response finishes, and when the workflow reaches a correct result. For a tool-using agent, include retrieval, external calls, retries and human review in that final measurement.
Set reasoning effort explicitly before comparing
There is a particularly important configuration detail. Inception documents four reasoning levels: instant, low, medium and high. Mercury 3 defaults to low when the setting is omitted, while Mercury 2.5 and Mercury 2 default to medium. The documentation associates higher effort with higher latency. Source: reasoning controls.
Consequently, our advice is to pin the setting in an evaluation rather than compare unnamed defaults. Record it alongside prompt version, output limit and serving route. Otherwise, a speed comparison can also become an unacknowledged comparison of deliberation settings.
A practical pilot decision
Start with one reversible task, such as drafting a support response for review. Use a fixed set of representative cases and decide what counts as an acceptable answer before running either model. Include ambiguous requests and cases that should be escalated rather than answered.
- Record first useful output and total completion time.
- Score correctness, tool arguments and escalation behavior.
- Calculate cost per accepted result, including retries.
- Compare against your current model using the same cases.
Advance only if the pilot improves the metric your users actually experience without weakening the acceptance standard. Mercury 3 is a timely candidate for that test, not a substitute for it.