Inception releases Mercury 3 with a larger context and changed reasoning default

Mercury 3 arrives for developers

Inception released Mercury 3 on October 19, making its latest language model available through its own API and through OpenRouter, Baseten and Vercel. The release expands the working context of its diffusion model family and targets applications that repeatedly call a model while answering a question or completing a task.

The accompanying API documentation reveals a consequential migration detail: requests that omit reasoning effort now select a different default. That can change behavior even when the rest of an application's request stays the same.

The transition also closes an entry point: Inception says Mercury 2 and Mercury 2.5 are no longer available to new users. Existing users have a limited migration window, while enterprise support follows their agreements.

Source 1Source 2Source 4Source 7

A larger text workspace

Mercury uses diffusion to generate tokens, the pieces of text a model processes, in parallel. For an application that pauses for several model responses, reducing the wait at each turn could shorten the overall task. That is an architectural rationale, not a measured improvement for every application.

The context window grows from 260,000 tokens in Mercury 2.5 to one million in Mercury 3. The listed maximum output also doubles to 131,072 tokens. Developers therefore have more room for source material and longer responses. The documentation lists text as Mercury 3's supported format; a larger context window does not add image or audio input.

Source 2Source 6

A default that can change a comparison

Inception's reasoning guide lists four effort settings: instant, low, medium and high. It describes a tradeoff between time spent deliberating and response latency, positioning high for complex reasoning and instant for the quickest interactions.

The migration detail is the default. Mercury 3 uses low when the developer omits reasoning_effort; Mercury 2.5 and Mercury 2 use medium. For eligible customers who still have legacy access, our analysis is that comparing untouched requests would mix a model change with an effort change. Setting the level explicitly makes that difference visible, although it does not guarantee equivalent reasoning or accuracy.

The quickstart itself uses medium. Its example should therefore not be read as demonstrating the new model's default behavior. Inception documents both its own SDKs and an OpenAI-compatible client configuration, giving existing applications a familiar integration route.

Source 3Source 4Source 7

What a database assistant actually does

Inception's SQL cookbook builds a sample shop database with customers, products and orders, reads its table structure, and supplies that structure in a tool definition. Mercury translates a question into a query; the application executes it, sends the results back, and requests an ordinary-language answer.

For a hypothetical question about pending orders, this pattern would use two model calls around a database operation. The model supplies the query and explanation, while the application provides access to the records. The example times those stages separately, illustrating why token-generation speed alone cannot describe the whole interaction.

There is also an implementation limit. The sample describes its SQL tool as read-only, but its function passes the generated query directly to SQLite's execute method without checking that it is a SELECT statement. Our reading is that the description expresses an intended restriction; it does not enforce one. A production adaptation needs database permissions or another enforced boundary. We inspected the published code but did not run it.

Source 5

Access and launch pricing

Inception's quickstart says new accounts receive 100 million free tokens without providing payment details. Developers create an account and API key, then add billing information to continue beyond the free allowance. This is documented self-service API access; we have not verified an individual account or regional eligibility.

Inception's models page limits Mercury 2 and 2.5 access to existing customers with usage before October 19, 2026, and says that access continues until November 15, 2026. Enterprise customers remain supported under their agreements. For other eligible legacy users, that gives migration and side-by-side checks a deadline; a newly created account cannot reproduce an old-versus-new comparison through those retired options. The API specification still lists their parameters, so its comparison table should not be read as a promise of new-user access.

Mercury 3's published launch rates are $0.10 per million input tokens, $0.01 per million cached input tokens and $0.35 per million output tokens. The regular rates are twice those amounts. The inspected pricing material does not give an end date for the launch discount, so the promotional rates should not be treated as a permanent price.

Source 2Source 3Source 7

What the speed claim establishes

Inception reports generation at 1,068 tokens per second on NVIDIA GPUs and claims stronger benchmark performance than GPT-6 Luna and Claude Haiku 5.5 across a range of tests. These are vendor claims, not measurements by AI Toolcast.

The SQL example requires two model calls and a database operation; generation speed alone cannot establish a complete workflow's latency or accuracy.

Source 1Source 5