# Mercury 3 API: Pricing, Million-Token Context and Diffusion Speed

Inception's new diffusion model expands context and cuts launch pricing, while existing Mercury users face a near-term migration deadline.

By Paras Madan | 2026-10-20


Inception launched Mercury 3 on October 19 with a million-token context window and an OpenAI-compatible API. Launch pricing is $0.10 per million input tokens and $0.35 per million output tokens. Its reported generation speed is the headline, but the migration deadline matters just as much for existing users.

The [official announcement](https://www.inceptionlabs.ai/blog/introducing-mercury-3) describes Mercury 3 as a diffusion-based language model. Inception reports 1,068 tokens per second on NVIDIA GPUs and says context has increased from Mercury 2.5's 260,000 tokens to one million. Those figures make it worth evaluating for applications where several model calls sit between a user's question and a useful answer.

## Mercury 3 API pricing

Inception's [model and pricing page](https://www.inceptionlabs.ai/models) separates standard rates from a 50% launch promotion. All prices below are per million tokens.

| Token type | Standard rate | Launch promotion |
| --- | --- | --- |
| Input | $0.20 | $0.10 |
| Cached input | $0.02 | $0.01 |
| Output | $0.70 | $0.35 |

The announcement does not give an end date for the promotion. Use the standard rates for a longer-term budget, and treat the discount as the current offer rather than a permanent price cut.

For example, a batch of requests totaling one million uncached input tokens and 100,000 output tokens would cost $0.135 during the promotion, or $0.27 at standard rates. That is API token cost alone. It excludes external tools, hosting and additional requests needed to correct or retry work.

Cached input costs one-tenth of normal input at either rate. That can matter for repeated instructions or shared context, but the saving depends on which requests actually qualify for caching. A price table cannot establish a workload's cache-hit rate.

The [quick-start guide](https://docs.inceptionlabs.ai/get-started) says new accounts receive 100 million free tokens without payment details. It describes an introductory allowance, not a recurring monthly allocation.

## How to use the Mercury 3 API

The model identifier is `mercury-3`. Inception documents the Chat Completions endpoint at `https://api.inceptionlabs.ai/v1/chat/completions`, with bearer-token authentication. Existing OpenAI SDK integrations can use `https://api.inceptionlabs.ai/v1` as their base URL.

The [model reference](https://docs.inceptionlabs.ai/get-started/models) lists text input and output, tool calling and structured outputs. Maximum output is 131,072 tokens, twice Mercury 2.5's 65,536. Reasoning effort supports `instant`, `low`, `medium` and `high`; the documented temperature range is 0.5 to 1.0.

The official example sets `reasoning_effort` to `medium` and uses `max_completion_tokens` to limit the response. Compatibility reduces integration work, but tool behavior and response formats still need testing before a production migration.

Inception also announces access through OpenRouter, Baseten and Vercel. Check the selected provider's model listing and billing before assuming that Inception's direct-API promotion applies through another service.

## Why diffusion speed matters for agents

Mercury refines text through a diffusion process. Its [streaming documentation](https://docs.inceptionlabs.ai/capabilities/streaming) describes ordinary streaming as blocks of refined text. The optional `diffusing=true` mode instead exposes intermediate versions: each chunk contains the full text so far, and a UI should replace its display rather than append every version. This is an integration detail worth checking if you want to show the refinement process.

For an agent, that speed is useful because delays accumulate. A browser agent might inspect a page, choose an action, read the result and decide what to do next. Reducing the model's contribution at each step can make the complete task feel more responsive, even when the agent takes the same number of steps.

However, **generation throughput is only one part of application latency**. A reported 1,068 tokens per second does not tell you how long the first answer takes, how much reasoning precedes it or how long a browser, database or search service spends returning information.

## What the benchmarks establish

Inception positions Mercury 3 against speed-oriented models such as GPT-6 Luna and Claude Haiku 5.5. Those comparisons are the vendor's claims; this article does not independently reproduce them.

The announcement also quotes TinyFish's evaluation on 50 live Mind2Web navigation tasks. TinyFish says Mercury 3 completed 50% in its production agent harness, using the same token volume and number of steps as a frontier-model comparison. The result is interesting because it examines an agent workflow, but 50 tasks in one harness cannot establish broad superiority across coding, reasoning or browser use.

A practical evaluation should therefore measure accepted results as well as elapsed time. A fast answer that requires another agent pass or substantial human repair can give back much of the initial saving. The million-token window similarly creates more room for information without proving that every distant detail will be used correctly.

## Mercury 2 and 2.5 migration deadline

The current model page says Mercury 2 and Mercury 2.5 are no longer available to new users. Existing customers with usage before October 19, 2026 retain access until November 15, 2026. Enterprise customers remain supported under their agreements.

That gives existing users a concrete reason to test Mercury 3 now. Keep a representative set of previous requests and compare output quality, tool behavior, cost and latency before moving production traffic. Where an application depends on a particular response format, include failures and edge cases rather than only successful demos.

Mercury 3's useful test is whether lower generation time creates enough room for a better product: another verification step, a faster interactive response or a more complete task within the same waiting time. That is a more meaningful outcome than making text appear faster on its own.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn model launches, API details and product tradeoffs into technical stories that developers and founders can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).