Mercury 3: Diffusion Text API, Pricing and Setup

Inception launched Mercury 3 on October 19, 2026. Its documented endpoint accepts text and supports tool calling and structured outputs, with up to 131,072 output tokens. Mercury Voice and Mercury Decide are separate products with different roles and billing; the Mercury 3 chat endpoint is the subject of this guide.

Diffusion generation is designed to reduce the time spent producing text. Inception reports 1,068 tokens per second on NVIDIA GPUs. Treat that as a publisher measurement: it does not specify the response time of your own long-context request, tool chain or network connection. Measure time to a useful answer as well as generation speed.

The direct Inception API lists a 50% launch discount as checked October 20. Prices below are USD per million tokens. The published offer does not confirm an end date, so budget ongoing work at the regular rate and use your current account quote for launch-period bills.

An assumed workload of one million fresh input tokens and 200,000 output tokens costs $0.17 at launch rates or $0.34 at regular rates. If 800,000 of those input tokens instead qualify as cached reads, the totals become $0.098 and $0.196. These arithmetic examples use the listed rates; they are not observed application bills or a promise of cache eligibility.

Sum usage across all calls before dividing spend by successfully completed tasks. An agent that retries a navigation step or repeatedly resends a document can use more tokens than a single answer suggests. Keep search, browser infrastructure and other tool charges in the same task ledger.

Create an Inception API account, set INCEPTION_API_KEY in your server environment and send a chat-completions request with model mercury-3. The example asks for a simple classification explanation, using medium reasoning and a bounded completion size. Executing it uses your account allowance or incurs API charges.

Read the returned answer and usage before adding the model to an agent. Tool calling returns a request for your application to handle; your code supplies the tools, permissions and execution loop. Check the provider’s schema support when migrating structured outputs rather than assuming every OpenAI parameter behaves identically.

The documented reasoning levels are instant, low, medium and high. Temperature accepts 0.5–1.0. Keep those settings explicit when comparing runs, especially if an existing integration sends a temperature or reasoning value outside that range.

For a support assistant, test whether the answer uses the right policy and escalates ambiguous cases. For a browser agent, record accepted task completion, steps, retries and elapsed time. Run short and long prompts separately: output generation throughput alone does not explain the delay before an answer begins or a tool finishes.

The one-million-token context is capacity for a request, not a requirement to fill it. Send the relevant material, then expand context only when it improves the result. Cache savings and a larger completion allowance are separate budget decisions.

Ellis Crosby’s October 20 Spring Prompt review tested inception/mercury-3 through OpenRouter, served by Inception, on October 19–20. In its speed-chess test, the author reports a median 0.42 seconds per move with reasoning switched off. At the API default, one move took 105 seconds after three invalid replies. The review labels its overall assessment provisional.

These are timed chess decisions, not a measurement of output tokens per second. The result makes reasoning settings and invalid-response handling worth testing for a deadline-driven application. Keep that small early evaluation separate from Inception’s generation-speed claim and from your own support, coding or browser tasks.

The launch names the Inception platform, OpenRouter, Baseten and Vercel as access channels. This page’s examples use the direct Inception endpoint and rates. A gateway can have its own routing, fees and availability; pin the intended provider when you compare quotes.

Mercury 3 is presented here as an API model. The checked release and model documentation provide no downloadable checkpoint for a local hardware estimate. Compare it with MiniMax M3.1 Flash on a complete task budget, or use a separately documented open-weight model when you need to run inference on your own hardware.