Gemini gains tools, tasks, and a simulated world
Google adds Tool Retrieval to Gemini, free agentic tasks, and Genie 4 for interactive simulation. The new features target both developers and general users, with tightly controlled costs and limits.
On November 19, 2026, Google introduced three distinct additions around Gemini: Tool Retrieval to manage large libraries of tools, free Task mode in certain countries, and Genie 4 to generate interactive worlds via the Gemini API. Together, these components expand the assistant's role from simple chat to task execution, purchasing, and real-time simulation. This move matters for developers as well as competitors of OpenAI and Anthropic, who charge for comparable capabilities.
Des outils moins lourds, mais pas magiques
Tool Retrieval, available in preview in the Gemini Interactions API, helps avoid a huge list of tools weighing down each request. The principle is simple: Gemini first keeps the name and description of functions, then loads the full schema only when the task requires it. Google explains that the classic approach quickly becomes cumbersome as soon as an agent has to juggle dozens or even hundreds of tools; input tokens grow, latency increases, and selecting the right tool degrades as the catalog expands.
The feature is only available through the Interactions API, which became generally available in June 2026. The old generateContent method does not support it. On the implementation side, Google offers a server mode with tool_search and defer_loading: true, or a client mode where the search is delegated to the developer's application. The two cannot be mixed in the same request, and only one tool_search is allowed. Convenient, but not exactly "plug and play" magic.
Pourquoi Google mise sur la récupération d’outils
The company says it recommends this approach starting from a library of about 20 to 30 tools. In its examples, the gain becomes clear when the collection is large: with 200 tools of 400 tokens each and only two tools loaded, Google mentions about 23,650 tokens versus 164,000 in static loading. However, a small catalog may lose the advantage, as the search adds a model call with all the context. In other words, efficiency has its conditions.
The technical framework remains precise. Retrieved schemas cannot be unloaded during a conversation, and the full tool table must be sent with each request, including with previous_interaction_id. Google also specifies that supported models are Gemini 4 Argon, Gemini 3.8 Flash, and Gemini 3.6 Flash. A request with defer_loading on an unsupported model returns a 400 error.
Gemini Task s’ouvre au public, avec des garde-fous
Google has also made Gemini Task free in several countries: United States, Canada, Australia, India, Japan, and South Korea. The agentic mode allows actions such as contacting businesses, checking slots, or initiating purchases, while in the United States the company adds automated phone calls, online payments, and access to password-protected sites. The challenge is both commercial and technical: facing OpenAI and Anthropic, Google tries to broaden usage without charging for entry.
The engine behind this ramp-up is Gemini 4 Hydrogen, a model presented as more efficient than Gemini 3.8 Flash in some internal tests, with data tasks up to 2.8 times faster. It can also handle multiple steps in parallel. Purchases go through Google Pay and the Agent Payments Protocol, with spending limits and mandatory user validation. Google says it is working with Mastercard and Visa for tokenized payments and plans to include American Express.
Genie 4 fait entrer la simulation dans l’API Gemini
Genie 4, Google DeepMind's new world model, is now accessible to developers via the Gemini API and Google AI Studio. The model, exposed as models/genie-4, generates a continuous world in 720p at 24 frames per second, with synchronized native audio. The user can move with WASD or arrow keys, then modify weather, lighting, or objects on the fly. Google presents it as its first world model intended for direct use by developers.
Advanced use cases include prototyping interactive experiences, visualizing spatial designs, and training robots or agents. Outputs carry an invisible SynthID watermark. On the transport side, Genie 4 uses WebRTC for browsers and WebSockets for servers, with ephemeral tokens to keep the main key off the client. The service requires a paid key, charges $1.50 per million input tokens, $9 per million output text tokens, and $17.50 per million output video tokens, with sessions limited to 120 seconds and hosting from the United States, Western Europe, and Japan.
Ce que Google cherche vraiment à vendre
Taken together, these launches outline a fairly clear strategy: make Gemini useful in real tasks while maintaining control over costs, security, and distribution. Tool Retrieval reduces the burden of tool-equipped agents, Gemini Task expands the assistant toward concrete action, and Genie 4 brings interactive simulation into developers’ toolkit. The real question remains, less glamorous but decisive: will reliability follow when these features move from testing to everyday use?
Key points
- Google announced Tool Retrieval on November 19, 2026.
- Gemini Task becomes free in six countries.
- Genie 4 enters the Gemini API on November 19, 2026.
- Google recommends retrieval starting from 20 to 30 tools.
- Gemini 4 Hydrogen is reportedly 2.8 times faster in some tests.
- Genie 4 limits sessions to 120 seconds.
By the numbers
- 26 tokens — indicative cost per deferred tool, according to Google's guide.
- 200 tools — comparative scenario cited by Google for Tool Retrieval.
- 400 tokens — size of each tool in the comparative example.
- 720p at 24 frames per second — video output of Genie 4.
- $1.50 / million tokens — entry price of Genie 4, according to Google.