Inception
Inception builds Mercury, a family of diffusion large language models that generate tokens in parallel instead of one at a time. Delivered through an OpenAI-compatible API, they target products where response latency is visible to users.
What is Inception?
Inception AI, Inc. describes itself as an AI research and product company building diffusion-based language models for production use. Its argument starts from a constraint: today's frontier models are autoregressive, emitting one token at a time from left to right, so latency and cost grow with every additional token and every additional step in a workflow. Diffusion works differently. Rather than committing token by token, it refines a whole output in parallel over a small number of coarse-to-fine steps, which allows revision during generation. The company's own image is that a model should behave less like a one-way typewriter and more like an editor.
That approach ships as Mercury, presented as the first commercially available family of diffusion large language models. Two are in production. Mercury 2 is the reasoning model, with a 128K context window, tool calling, structured outputs and adjustable reasoning effort. Mercury Edit 2 is a small code-focused model with 32K windows for fill-in-the-middle autocomplete and next-edit prediction. Both accept text only, with no image input and no image generation. Both are priced identically: 0.25 USD per million input tokens, 0.025 USD for cached input, 0.75 USD per million output tokens. Mercury 1 remains supported for existing customers.
The claimed numbers are the pitch, and they belong to the vendor: under 300 ms to first token, five to seven times higher throughput, 70% lower cost per task, and over 1,000 tokens per second on commercial NVIDIA GPUs, at a quality the site aligns with speed-optimised autoregressive models. Access is deliberately unremarkable: the API is OpenAI-compatible and described as a drop-in replacement, with official Python and TypeScript SDKs and support through AISuite, LiteLLM and LangChain. Editor and voice platforms — Cursor, Zed, Kilo Code, Roo Code, OpenCode, OpenClaw, ElevenLabs, Vapi — each have a setup page, and the models are also sold through AWS Bedrock, Azure Foundry, OpenRouter and Models.dev.
Enterprise options cover no-logging and no-retention modes, private endpoints and VPC connectivity, dedicated capacity and custom terms, with more than 99.5% uptime and bespoke SLAs. The founding team comes from Stanford, UCLA and Cornell, and lists twenty publications between 2019 and 2025.
What it does
- Generate text and reasoning through the v1/chat/completions endpoint with the mercury-2 model
- Autocomplete code inline through fill-in-the-middle on v1/fim/completions with mercury-edit-2
- Predict a developer's next edit through the v1/edit/completions endpoint
- Call external functions and orchestrate multi-step agents with tool calling
- Force answers into a JSON schema using structured outputs
- Dial reasoning effort down to an instant mode for near real-time voice turns
- Run the same models through AWS Bedrock, Azure Foundry or a model router instead of the direct API
When to use Inception / When not to
A quick filter to help you decide if Inception is the right fit.
When to use Inception
- Developers embedding a language model in a product where response time is felt directly by the end user
- Teams building coding tools — inline autocomplete, next-edit suggestions, in-flow chat — with documented setups for Cursor, Zed, Kilo Code, Roo Code and OpenCode
- Builders of real-time voice experiences such as customer support, live translation or immersive gaming, through ElevenLabs or Vapi
- Engineers assembling tool-calling agents that need structured outputs and many fast turns rather than one long answer
- Enterprise buyers who prefer to procure through AWS Bedrock or Azure Foundry instead of signing a new vendor contract
When not to use Inception
- Anyone who needs image or vision input: both Mercury models accept text only, and neither generates images
- Non-technical users, since everything outside the Mercury Chat sandbox requires an API key and code
- Mobile-first teams expecting a native app, as there is no iOS or Android release
- Regulated organisations that need published assurance: no SOC 2, ISO 27001, HIPAA or GDPR statement appears anywhere on the site
- Buyers who want a fixed monthly subscription, since billing is metered per million tokens with no flat plan to cap a budget
How to use Inception
A typical end-to-end flow, from setup to results.
- Try the models without writing code first, in the Mercury Chat sandbox at chat.inceptionlabs.ai
- Open an account on platform.inceptionlabs.ai: every new account is credited with 100 million free tokens and no payment details are required
- Create an API key under API Keys in the dashboard
- Export it as the INCEPTION_API_KEY environment variable
- Send a first request to https://api.inceptionlabs.ai/v1/chat/completions with model mercury-2, in cURL, Python or TypeScript
- Or install the official SDK, the inceptionai package in both Python and TypeScript, which reads the key from the environment
- For coding work, call v1/fim/completions or v1/edit/completions with mercury-edit-2 instead
- Switch on tool calling, structured outputs, streaming or instant mode through request parameters as needed
- Paste the same key into a third-party client such as Cursor, Zed, Kilo Code, Roo Code, OpenCode, OpenClaw, ElevenLabs, Vapi or LangChain
- When the free tokens run low, add a payment method under Billing, and email support with your use case and expected volume to raise quotas
Pros & Cons
Pros
- Speed and cost are the whole proposition, and the figures are specific: under 300 ms to first token, five to seven times the throughput, 70% less per task
- Public, uniform pricing across both models, with cached input ten times cheaper than fresh input
- 100 million free tokens on sign-up, with no card required
- OpenAI compatibility turns migration into a configuration change, and third-party integrations are documented one by one
- The training opt-out is named and locatable, the Improve the model for everyone switch, which few vendors document at all
- Enterprise terms advertise no training on customer data, alongside no-logging modes and VPC connectivity
- Sold through AWS Bedrock and Azure Foundry, so procurement can run on an existing cloud contract
Cons
- No published security or compliance assurance: SOC 2, ISO 27001 and HIPAA appear nowhere on the site
- No GDPR mention at all, no Article 27 representative and no data processing agreement
- No hosting location disclosed, neither country nor region
- Text input only, with no image input and no image generation
- Metered billing with no flat plan makes a budget hard to cap, especially for an agent running in a loop
- Every performance and cost figure is the vendor's own, and the Models page carries two different comparison sets across its rendering variants
- The terms take a perpetual, transferable licence on submissions with model training among the stated purposes, leaving the opt-out to the user
Pricing & Plans
A permanent free tier is available: the Free plan grants access to all models with 100 million tokens included and no payment details required. Beyond it there is no fixed subscription. The paid entry point is metered usage at 0.25 USD per million input tokens, 0.025 USD per million cached input tokens and 0.75 USD per million output tokens, identical for Mercury 2 and Mercury Edit 2. Enterprise pricing is volume-based and quoted on request.
- access to all models and 100 million free tokens
- with no payment details required
- usage-based pricing
- generous rate limits and priority support
- the documentation calls this tier Pay As You Go
- custom rate limits
- SLA guarantees
- security and privacy commitments
- volume-based pricing
- through Contact Sales
- 0.25 USD per million input tokens
- 0.025 USD per million cached input tokens
- 0.75 USD per million output tokens
Data, GDPR & hosting
A consolidated view of how Inception handles your data.
GDPR overview
There is no GDPR implementation to report: the words GDPR and CCPA appear nowhere across the thirteen pages collected. No Article 27 representative, no data protection officer, no legal basis, no transfer mechanism and no data processing agreement is mentioned. The privacy policy, effective 4 April 2025, rests on United States law and offers state-level rights instead: California residents may object to disclosure for third-party direct marketing, which Inception says it does not practise, and the company states it does not sell personal data as Nevada defines it. Do Not Track signals are not honoured. Privacy questions go to legal@inceptionlabs.ai, while privacy@inceptionlabs.ai handles matters involving children. A European buyer therefore has nothing published to rely on and must obtain commitments contractually.
Who owns the data?
Submitting content grants Inception a worldwide, non-exclusive, perpetual, royalty-free, sublicensable and transferable licence to use, edit, modify, truncate, aggregate and otherwise fully exploit those submissions for the purpose they were sent. The terms name four uses: running the service, complying with law, enforcing policies, and training the models. Users keep their own ownership rights and may license their material elsewhere; the service itself remains Inception's property. Personal data goes only to service providers — Stripe, Inc. for payments and Google LLC for analytics are the two named — and may pass to an acquirer in a merger, acquisition or bankruptcy. Enterprise contracts reverse the training position entirely.
Reuse rights
Users receive a worldwide, non-exclusive, non-sublicensable and non-transferable licence to use Content solely for operating the service. Reproduction, modification, distribution or storage for any other purpose requires prior written permission, and the presence of a copy or download button does not lift that restriction. The terms are silent on who owns model outputs: no clause assigns them to the user, while liability for errors in outputs is pushed onto the user. Submissions remain the user's own property, and further licences may be granted on them. Anyone planning commercial reuse of generated text should settle the point contractually rather than infer it from these terms.
Data retention & training
Hosting summary
No hosting jurisdiction is published. Neither the privacy policy, nor the terms of use, nor the Enterprise page names a country or region where data is stored. Two subprocessors are named: Stripe, Inc. for payments and Google LLC for analytics. Other providers are described only by category — hosting, technology and communications, analytics, security and fraud prevention, support, payment processing. Security is stated in general terms, with appropriate physical, technical, organisational and administrative measures and the usual caveat that no transmission is entirely safe. Enterprise deployments offer more: prompts and outputs treated as customer data, configurable retention and caching, no-logging or no-retention modes where applicable, private endpoints or VPC connectivity, and dedicated capacity. AWS Bedrock and Azure Foundry are purchasing channels rather than declared locations, and their regions are not specified. One point not to confuse: the site's public IP resolves to an anycast CDN node in Amsterdam on Amazon's network, which describes delivery of the marketing site, not where inference runs or data rests. Personal data may also pass to an acquirer in a merger, acquisition or bankruptcy.
Where Inception works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Inception.
- The licence on submissions is perpetual, transferable and sublicensable, with model training among its stated purposes: a prompt can feed a commercial model unless the user turns the switch off
- That opt-out lives in User Settings on the API Platform, so from the user's point of view it is off by default and easy never to find
- The terms never say who owns model outputs, while pushing liability for their errors onto the user
- With no certification and no data processing agreement published, a regulated organisation has nothing to rely on but a negotiated contract
- Speed is a cognitive trap: an answer in 300 ms invites you not to reread it, and the quality target is a speed-optimised model rather than a frontier one
- Metered billing without a ceiling means an agent stuck in a loop can spend far beyond the intended budget
- Disputes go to individual arbitration in San Francisco, with class actions and jury trial waived; opting out requires a posted letter within thirty days
Setup & Integrations
Technical difficulty
Developer-level. Outside the Mercury Chat sandbox everything runs on an API key, so there is no no-code path. The documented start is three steps — create an account, create a key, send a request — and no card is needed. Teams already calling OpenAI change a base URL, a key and a model name. Official Python and TypeScript SDKs read the key from the environment, and editor integrations are a copy-paste. The real work is elsewhere: the code endpoints, fill-in-the-middle and next edit, do not behave like chat, and exponential backoff on 429 and 503 is expected.
Deployment
Integrations
Supported languages
Behind Inception
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Inception.
Frequently asked questions
What is a diffusion LLM, and how does it differ from a conventional model?
Which models can I use?
What does it cost?
Can I try it for free?
Will it work with code written for OpenAI?
Do the models accept images?
Is my data used to train the models?
Which integrations are documented?
Is there a minimum age?
Is uptime guaranteed?
Should you pick Inception?
Inception is one of the few AI companies whose differentiator is architectural rather than incremental: Mercury is diffusion applied to text, not a faster variant of the same autoregressive recipe. That makes it worth an hour from anyone whose product is judged on responsiveness — streaming code completions, live voice turns, agents that take many short steps instead of one long one. The evidence behind the claim is unusually checkable for a young company: twenty dated publications, the Mercury paper on arXiv, three identified academic founders, and a 50 million USD seed round.
Trying it costs nothing and commits nothing. A hundred million free tokens arrive with the account, no card is asked for, the API answers OpenAI-shaped requests, and editor integrations are documented one page at a time. If the latency figures hold on your own inputs, the economics follow.
Production is a different question, and the gaps sit on the compliance side rather than the technical one. Nothing on the site names a hosting country, a certification, a data processing agreement or the GDPR, and the terms take a broad, perpetual licence on submissions with model training among its purposes. The opt-out exists and is named, but the user has to go and switch it off. The Enterprise page promises the opposite regime — no training on customer data, no-logging modes, private networking — which means the real answer depends on the contract rather than on the published policy.
Read the performance numbers as the vendor's own, since no third-party benchmark appears on the site, and note that only two models are in production, text input only. For a latency-bound product built by a team willing to negotiate terms, that is a reasonable trade.
- Choosing a selection results in a full page refresh.
- Opens in a new window.