Google Gemini 3.5 release: action models and tool-use API

Google unveiled Gemini 3.5 at its I/O developer conference, releasing a family of models designed to pair advanced reasoning with direct action capabilities. The rollout includes variants tuned for tool use, multimodal inputs, and larger context windows, aimed at developers and enterprise customers.

Google positions Gemini 3.5 as the next step in the Gemini line, describing the series as combining what it calls frontier intelligence with 'action' features that let models control external tools and execute workflows. The company highlighted improvements in the models' ability to compose multi-step tasks, call APIs, and operate over longer documents and codebases. Google also emphasized safety and guardrails for models that can act on behalf of users.

What’s new in Gemini 3.5

Gemini 3.5 is presented as a family rather than a single checkpoint. Google describes distinct variants that prioritize different trade-offs between latency, capability, and tool integration. Key points announced by Google include:

• Action-enabled model variants, designed to call external APIs, run code snippets, and interact with developer tools.
• Multimodal support, with models able to accept images alongside text and respond using a unified reasoning pipeline.
• Expanded context windows intended to let the models work over longer documents, code repositories, or multi-message conversations.
• Updated developer interfaces and SDKs that document how to wire model outputs to third-party tools and cloud services.
• Emphasis on safety features and usage controls for models that can take actions, including monitoring and policy hooks.

Google framed these features as intended to reduce manual glue code for integrating models with existing systems. The company signaled that some variants will target latency-sensitive applications while others prioritize deep reasoning and tool orchestration.

Availability and developer access

Google said Gemini 3.5 will be accessible to developers through its API surface and cloud offerings, with staged availability across developer previews and enterprise tiers. Pricing, rate limits, and exact deployment footprints were characterized as dependent on variant and usage patterns, with more details to appear in product documentation and platform dashboards.

Integrations showcased by Google included sample agent flows that link model outputs to API calls, code execution, and queryable knowledge stores. Google is positioning partner tooling and enterprise controls as part of the offering for customers that need stricter governance and audit trails when models take action.

Early developer materials focus on examples for automating data workflows, generating and running code, and using multimodal inputs to annotate or summarize visual content. Google reiterated that safety and alignment work continues, especially around systems that can perform external operations on behalf of users.

Why it matters

Combining stronger reasoning with built-in action capabilities shifts some of the engineering work for agentic systems from glue code into the model and platform. That can accelerate developer integration for productive workflows, while also raising governance and safety demands for organizations that allow models to make external calls or execute code. Enterprises and platform builders that need controlled automation will be the earliest adopters and will shape how action-enabled models are governed in production.