Inception Platform — API docs (Mercury diffusion LLMs)
First-party developer documentation for inception-labs's API. It sells three models — Mercury 2.5, Mercury 2 and Mercury Edit 2 — described as diffusion LLMs delivering "frontier-quality reasoning and code generation at a fraction of the latency." The API is OpenAI-compatible, so the whole product is a drop-in swap for an existing client.
This is the spoke's first commercial text-model provider that is not a transformer decoder shop. Diffusion appears elsewhere here only in image and media models (mage-flow-turbo); this is diffusion sold for chat, reasoning, tool calling and code.
Models and prices
| Model | Input /1M | Cached input | Output /1M | Context | Max output | Endpoints |
|---|---|---|---|---|---|---|
| Mercury 2.5 | $0.20, discounted to $0.04 | $0.02 → $0.004 | $0.75 → $0.15 | 260K | 65,536 | chat |
| Mercury 2 | $0.25 | $0.025 | $0.75 | 128K | 50,000 | chat |
| Mercury Edit 2 | $0.25 | $0.025 | $0.75 | 32K (FIM and NextEdit) | 8,192 | FIM, edit |
The headline row is a promotion — an "80% off Mercury 2.5" banner runs across the docs — so $0.04 / $0.15 is a sale price, not a rate card. At list ($0.20/$0.75) Mercury 2.5 sits in llm-api-pricing's budget tier beside Mistral Small and GPT-4.1 Nano; at the promotional price it undercuts everything the spoke holds. Cached input is a flat 10× discount on input across all three models, which is a cleaner rule than most providers publish.
Text only. No vision, no image input, no image generation — stated plainly in the FAQ, and worth noting for a company whose architecture comes from image models.
What diffusion changes at the API
- Streaming is block-wise, not token-wise. The docs define a block as "a chunk of refined text the model emits as it decodes, rather than one token at a time" — the honest surface consequence of denoising a span in parallel instead of running an autoregressive loop.
- A "diffusing" mode streams the intermediate noisy states, so a client can show text being refined into the answer. No other provider here exposes anything comparable, because no other provider has intermediate states worth showing.
- Everything else is deliberately ordinary: OpenAI-compatible chat completions, tool calling,
structured outputs via JSON schema, prefix caching with
cached_tokensin the usage object, reasoning tokens counted at the output rate.
reasoning_effort has four levels — instant, low, medium (default), high — described by
latency and use case (instant for voice and realtime turns; high for multi-step reasoning). The
docs give no quality or benchmark difference between them, only a latency ordering.
Code as the second product
Mercury Edit 2 is a code model with two purpose-built endpoints rather than chat: FIM (fill-in-the-middle, prefix plus optional suffix, for inline completion) and Next Edit, which predicts a developer's next edit from recent snippets, file content and edit history. Neither supports tool calling. The integration list is IDE-shaped — Cursor, Zed's edit predictions, VS Code, Kilo Code, OpenCode, Claude Code via a LiteLLM bridge, plus LangChain, ElevenLabs and Vapi for voice.
Latency is the whole pitch, and it is aimed at the two markets where autoregressive decode hurts most: inline code completion and realtime voice.
Access
Free tier at 1,000 requests / 1M input tokens / 100k output tokens per minute; Pay As You Go triples each; Enterprise is custom. Every new account gets a one-time 100M free tokens shared across models. No dedicated MCP endpoint — OpenAI-compatible tool calling is offered as the substitute.
The docs publish an llms.txt index and state that the site does not block crawlers or agent
traffic, a small first-party data point for ../search-marketing-wiki's crawl-economy thread: a
vendor courting agent readers while publishers meter them.
Why T1, and what is missing
First-party product documentation — the primary record of what is sold, at what price, with what
limits. What it does not contain is any evidence for the claims that matter: no benchmark, no
tokens-per-second figure, no latency measurement, no comparison against any named model. "Fastest"
and "frontier-quality" are asserted. The docs carry no dates, so published: is left off, and
freshness: volatile reflects a live price list under a promotion.
Connections
- inception-labs — the provider; llm-provider — the landscape
- llm-api-pricing — where these prices land, at list and on sale
- cross-spoke: the mechanism — block-wise decoding, denoising instead of an autoregressive loop —
is
../llm-inference-wiki's subject, and it has no page on diffusion decoding; this is the seam - cross-spoke: the IDE and agent integrations are
../agentic-tooling-wiki's