The post, on the Google for Developers blog, accompanies Gemini 3’s availability to developers through the Gemini API. It describes Gemini 3 as Google’s most intelligent model and says that several API updates were rolled out to support its reasoning, autonomous coding, multimodal understanding and agentic capabilities. The stated aim of the changes is to give developers more control over how the model reasons, how it processes media, and how it interacts with the outside world.
The first half lists what is new in the API; the second gives recommendations for getting the best results from Gemini 3 Pro, which the post calls Google’s most advanced model for agentic coding, and points to a System Instructions template that Google’s research team created for it and that the post says improved performance on several agentic benchmarks.
Key Points
- A new
thinking_levelparameter controls the maximum depth of the model’s thinking before it responds. Gemini 3 treats the levels as relative guidelines for reasoning rather than strict token guarantees; the post suggests “high” for complex tasks such as strategic business analysis or scanning code for vulnerabilities, and “low” for latency- and cost-sensitive work such as structured data extraction and summarization (see Reasoning Effort). - A
media_resolutionparameter sets how many tokens are used for image, video and document inputs, per media part or globally, trading visual fidelity against token usage and latency; if it is left unspecified, the model uses defaults based on the media type. - Starting with Gemini 3, the API enforces the return of Thought signatures, encrypted representations of the model’s internal thought process that keep its chain of reasoning intact across a conversation. Function calling validates them strictly on the current turn and image generation or editing for all model parts, both answering a missing signature with a 400 error; text and chat generation does not enforce them strictly, but omitting them degrades reasoning and answer quality. The official SDKs with standard chat history handle them automatically.
- Gemini’s hosted tools — specifically Grounding with Google Search and URL context — can now be combined with structured outputs, which the post presents as useful for agents that fetch live information from the web and extract it into a precise JSON format.
- Pricing for Grounding with Google Search moves from a flat per-prompt rate to a usage-based rate per search query.
- The prompting recommendations for Gemini 3 Pro are Google’s own: keep temperature at its default of 1.0; keep a uniform prompt structure and define ambiguous terms; ask explicitly for a more conversational answer, since the model is by default less verbose; treat text, images, audio and video as equal-class inputs; put behavioral constraints and role definitions in the system instruction or at the top of the prompt; and for long contexts, put the specific instructions at the end, after the data.
Context
The post is a vendor announcement written for developers adopting the new model, and each change it describes is framed around agentic use: thought signatures as critical for multi-step agentic workflows, grounding combined with structured outputs as a building block for agents, and the pricing change as better supporting dynamic agentic workflows. It names vibe coding among the uses it reports wide excitement for with Gemini 3 Pro, alongside zero-shot generation, mathematical problem solving and complex multimodal understanding. It directs readers to the Gemini 3 documentation and developer guide for implementation details.