Mercury 2 at a glance

Mercury 2 is Inception's current general-purpose reasoning diffusion language model. It is an API model, not a complete phone system: a voice deployment still needs speech recognition, text-to-speech, telephony, orchestration, monitoring and human escalation.

Diffusion generation works by refining a span of noisy text over repeated denoising passes, allowing many token positions to be processed together. That differs from conventional left-to-right autoregressive decoding and is the basis for Inception's low-latency claims.

Inception reports more than 1,000 decoded tokens per second on standard NVIDIA H100 GPUs and roughly 300 milliseconds for a 300-token reasoning trace. Those figures are vendor measurements, not guaranteed end-to-end response times.

The model exposes instant, low, medium and high reasoning-effort settings. Instant prioritizes responsiveness and is documented for turns that do not need tool calls; medium is the official default for most applications.

The OpenAI-compatible chat-completions endpoint supports text, a 128K context window, a 50,000-token maximum output, tool calling and structured outputs. Streaming returns blocks, while an optional diffusing mode exposes changing intermediate denoising text and should not be shown as a stable final answer.

Inception highlights realtime voice, customer support, enterprise search, workflow subagents and rapid coding iteration. Its voice cost example estimates only the model layer and explicitly excludes transcription, speech synthesis and telephony.

Current list prices are $0.25 per million input tokens, $0.025 per million cached-input tokens and $0.75 per million output tokens. Inception's docs currently advertise 100 million free tokens for a new account, although one marketing-page line still says 10 million; confirm the actual grant and expiration in the platform.

A low model-token price does not equal a low total voice-call price. Resent conversation history, reasoning tokens, retries, tool calls, retrieval, speech services, phone minutes, logging and human transfers can dominate the real bill.

Inception's terms prohibit misinformation, political campaign content and high-stakes decision-making involving individual rights. They also say the service is for internal personal use and not on behalf of a third party, language that production or customer-facing buyers should reconcile with Inception in writing.

Prompts and other user submissions may be used to train Inception's models unless the account setting called Improve the model for everyone is turned off. The privacy policy also covers personal data in prompts and uploads and provides only purpose-based, not fixed, retention language.

Fast reasoning does not remove hallucinations, prompt injection, tool mistakes or policy failures. Voice agents need deterministic authorization, least-privilege tools, confirmation before consequential actions, transcript review, disclosure that the caller is speaking with AI and an immediate human handoff path.