Inception API

Mercury 2.5 is a diffusion language model from Inception Labs. Unlike autoregressive models that emit one token at a time, it generates and refines tokens in parallel. Inception reports 1,107 tokens per second on standard GPUs, and says quality is comparable to cost-optimized models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. Reported benchmark figures include 79% on GPQA Diamond and 77% on IFBench. Independent benchmark coverage was not published at release, so those numbers are Inception's own. Through the API it supports tunable reasoning, parallel tool calls, and schema-aligned JSON output. With a 260K context window and 65,536 max output tokens, it targets latency-sensitive work such as search and RAG pipelines that fire many calls per request, voice agents, and coding-assistant tasks like context compaction and routing.

Mercury 2.5 Preview is a diffusion-based reasoning language model from Inception Labs, refining tokens in parallel rather than generating them one at a time. It preceded the general release of Mercury 2.5 and followed Mercury 2, adding tunable reasoning levels, parallel tool calls, and schema-aligned JSON output. Inception reports throughput of 1,107 tokens per second on standard GPUs and a 10+ point jump in intelligence over Mercury 2. It says quality is comparable to cost-optimized models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. With a 260K context window and 65,536 max output tokens, it targets production workloads where latency compounds, such as search agents, voice pipelines, and coding subagents. It is available through OpenRouter with an OpenAI-compatible API, priced at $0.04 per million input tokens and $0.15 per million output tokens.

Mercury 2 is a diffusion-based reasoning language model from Inception Labs that refines all tokens in parallel rather than generating them sequentially, achieving over 1,000 tokens per second — roughly 5x faster than speed-optimized competitors like Claude Haiku and GPT-5 Mini at comparable quality. On reasoning benchmarks, Mercury 2 scores 91.1 on AIME 2025 and 73.6 on GPQA. It also placed second on the Copilot Arena leaderboard for quality while ranking first for speed overall. With a 128K context window, it is purpose-built for latency-sensitive applications — real-time assistants, high-throughput pipelines, and cost-conscious production workloads where reasoning capability matters.

Mercury Coder is a code-specialized diffusion language model from Inception Labs, built on the same parallel token refinement architecture as Mercury. It is available in Mini and Small sizes. On fill-in-the-middle tasks, Mercury Coder Small scored 84.8% average accuracy, exceeding Codestral 2501 (82.5%). On MultiPL-E, it reaches 82.0% in C++, 83.9% in JavaScript, and 82.6% in TypeScript. In Copilot Arena human evaluations, Mercury Coder Mini ranked second in user preference with an average latency of just 25 milliseconds. It is the go-to choice for real-time code completion, autocomplete, and apply-edit workflows where both speed and accuracy are critical.

Mercury is the world's first commercial-scale diffusion large language model from Inception Labs. It generates text through iterative parallel refinement rather than sequential token prediction, enabling dramatically higher throughput without sacrificing output quality. It matches the performance of frontier speed-optimized models such as GPT-4o Mini and Gemini 1.5 Flash across knowledge, coding, instruction-following, and math benchmarks, while running up to 10x faster. It is OpenAI API-compatible for straightforward integration. Mercury is well-suited for API use cases that demand high concurrency, fast response times, or cost efficiency — including chat, summarization, and general-purpose text generation at scale.