Managed RAG and Genie 4: two bets on AI operated by services
Perplexity and Google illustrate two distinct trends: outsourcing document search for enterprise AIs, and exposing Genie 4 via API for short interactive worlds. In both cases, the challenge is no longer just the model, but the operational aspect.
On November 19, 2026, Perplexity and Google published two announcements that say a lot about how AI is now marketed to businesses and developers. On one side, Perplexity describes RAG as a service, a way to outsource context search for enterprise AIs. On the other, Google opens Genie 4 via the Gemini API and Google AI Studio, with interactive sessions limited to 120 seconds. In both cases, the focus is no longer just on generation, but on how it is operated, billed, and maintained.
AI no longer just relies on a model, it is also rented à la carte
The guide published by Perplexity presents RAG as a service, or RaaS, as a model where a provider handles all or part of the retrieval-augmented generation components. RAG, as a reminder, allows an AI to fetch information outside its training parameters before responding. Here, the shift is clear: vector databases, embedding models, retrievers, ingestion, storage, security, and optimization can be transferred to the provider.
Perplexity distinguishes three setups. Local RAG remains entirely on the company’s servers. A cloud system built internally relies on external APIs, but orchestration remains in-house. With RaaS, the provider manages the bulk and the client retains only part of the plumbing. The result is less spectacular than a slogan, but more concrete for IT teams: less infrastructure to manage, more dependency on the provider.
Why companies are taking a closer look
The guide cites several use cases: internal knowledge management, customer support, sales assistance, access to customer data, and product-oriented chatbots. An internal assistant can query Google Drive, SharePoint, or Confluence. A sales tool can leverage support tickets, competitive research, or call recordings. On the client side, the same pattern is used to manage subscriptions, returns, or personal data changes.
Perplexity highlights faster production deployment, less maintenance, better load absorption capacity, automatic content synchronization, and integration with enterprise platforms. But the company also reminds that service and plan limits apply. The debate remains very pragmatic: building internally can avoid some token and storage costs, but hardware and operations may offset the advantage. In short, the cloud hasn’t disappeared, it just changed aprons.
Google turns Genie 4 into a tool for short sessions
Meanwhile, Google has made Genie 4 accessible via the Gemini API and Google AI Studio. The model produces interactive worlds in continuous 720p video at 24 frames per second, with synchronized native audio. Developers can start from text, a reference image, or both, then control the scene with WASD keys, arrows, or textual events. Google presents the tool as useful for prototyping interactive experiences, visualizing spatial concepts, or training robots and AI agents.
The promise remains tightly framed. A Genie 4 session lasts a maximum of 120 seconds, operates on a sliding time window, and does not maintain a persistent world. The guide notes that visual stability may degrade after 60 to 90 seconds, and the system does not provide a lasting map of the environment. For robotic uses, this is significant: a setting that forgets a door after a U-turn is not yet a very polished training ground.
What Google charges, and what the API does not allow
Genie 4 is not available on the free plan. Google states a price of $1.50 per million input tokens of text or reference image, $9 per million output text tokens, and $17.50 per million output video tokens with stereo audio. The company also specifies that batch mode, Flex inference, and Priority inference are not supported. Sessions run via WebRTC or WebSocket, with an ephemeral token on the browser side if needed.
Session end control is explicit: closing a tab does not necessarily stop generation on the server side. Developers must call complete() to end the session and billing, or cancel() to interrupt execution. Google also enforces quotas per paid project, for example 2 active sessions and 50 requests per day at tier 1, 5 sessions and 250 requests at tier 2, then 10 sessions and 1,000 requests at tier 3.
Two products, one same industrialization logic
Both announcements show the same direction: useful AI is moving towards operated, monitored, and usage-based billed services. Perplexity offers to handle part of the document chain for companies. Google pushes Genie 4 towards brief, controlled sessions, calibrated for prototyping rather than persistent worlds. The technical approaches differ, but the trade-off is identical: outsource what costs time, infrastructure, and complexity, while keeping an eye on provider dependency and platform limits.
Within this framework, companies are not forced to choose a single camp. Perplexity also points out that the same organization can keep simple systems internally and delegate the most complex cases. This is probably where the next market stage is played out: no longer “buying AI,” but composing a chain of services where each link is subcontracted, kept, or replaced according to real cost.
Key points
- Perplexity talks about RAG as a service on November 19, 2026.
- Google limits Genie 4 to 120 seconds per session.
- Genie 4 charges $17.50 per million video tokens.
- Perplexity targets 50,000 organizations on Enterprise.
- Google enforces up to 10 active sessions at tier 3.
- Both offers shift operational tasks to the provider.
By the numbers
- 120 seconds — maximum duration of a Genie 4 session, according to Google, 2026.
- 720p at 24 fps — continuous video output of Genie 4, according to Google, 2026.
- $1.50 / million tokens — price for input text or reference image, Google, 2026.
- $17.50 / million tokens — price for video output with stereo audio, Google, 2026.
- 50,000 organizations — companies claiming to trust Perplexity Enterprise, enterprise page, 2026.