Gemini 3, our most clever mannequin is on the market for builders to construct with by way of the Gemini API. To help its state-of-the-art reasoning, autonomous coding and multimodal understanding and highly effective agentic capabilities, we’ve rolled out a number of updates to the Gemini API. These adjustments are designed to offer you extra management over how the mannequin causes, the way it processes media, and the way it interacts with the surface world.
Here’s what is new within the Gemini API for Gemini 3
- Simplified parameters for considering management: Beginning Gemini 3 onwards, we’re introducing a brand new parameter known as thinking_level to regulate the utmost depth of the mannequin’s considering course of earlier than it produces a response. Gemini 3 treats these ranges as relative pointers for reasoningrather than strict token ensures. The thinking_level parameter means that you can regulate the depth of the mannequin’s inner reasoning. You possibly can set it to “excessive” for complicated duties that require optimum considering (e.g. strategic enterprise evaluation, or scanning code for vulnerabilities). You possibly can set it to “low” for latency and cost-sensitive purposes akin to structured information extraction, summarization and so forth. Learn extra right here
- Granular management over multimodal imaginative and prescient processing: The media_resolution parameter permits you to configure what number of tokens are used for picture, video and doc inputs , permitting you to stability visible constancy with token utilization. The decision could be set utilizing media_resolution_low, media_resolution_medium, or media_resolution_high per particular person media half or globally. If unspecified, the mannequin makes use of optimum defaults based mostly on the media sort. Increased resolutions enhance the mannequin’s capacity to learn high quality textual content or determine small particulars, however enhance token utilization and latency.
- Thought signatures to enhance operate calling and picture era efficiency: Beginning with Gemini 3, we’re implementing the return of “Thought Signatures.“. These are encrypted representations of the mannequin’s inner thought course of. By passing these signatures again to the mannequin in subsequent API calls, you make sure that Gemini 3 maintains its chain of reasoning throughout a dialog. That is crucial for complicated, multi-step agentic workflows the place preserving the “why” behind a choice is simply as vital as the choice itself. Should you use the official SDKs and normal chat historical past, thought signatures are dealt with robotically for you. On the API the validations work as follows
- Operate Calling has strict validation on the “present flip”. Lacking signatures will lead to a 400 error. For understanding of how signatures present up for numerous operate calling situations please learn right here
- For Textual content/chat era, validation isn’t strictly enforced, however omitting signatures will degrade the mannequin’s reasoning and reply high quality.
- Picture era/enhancing has strict validation for all mannequin elements together with a thoughtSignature. Lacking signatures will lead to a 400 error.
- Grounding and URL Context with Structured Outputs: Now you can mix Gemini hosted instruments, particularly Grounding with Google Search and URL context with structured outputs. That is particularly highly effective for constructing brokers that must fetch reside data from the net or particular webpages and extract that information right into a exact JSON format for downstream duties.
- Updates to Grounding with Google Search pricing: To raised help dynamic agentic workflows, we’re transitioning our pricing mannequin from a flat price (US$35/1k prompts) to a extra granular, usage-based price of US$14 per 1,000 search queries.
Finest practices for utilizing Gemini 3 Professional via our APIs
We’ve got seen broad pleasure for Gemini 3 Professional particularly with vibe coding, zero-shot era, mathematical drawback fixing, complicated multimodal understanding challenges and a wide range of different use circumstances. So as to get one of the best outcomes whereas pushing the boundaries of Gemini 3. Extra particulars right here.
- Temperature: We strongly advocate maintaining the temperature parameter at its default worth of 1.0
- Consistency & Outlined Parameters: Preserve a uniform construction all through your prompts (e.g., standardized XML tags) and explicitly outline ambiguous phrases.
- Output Verbosity: By default, Gemini 3 is much less verbose and prefers offering direct, environment friendly solutions. Should you require a extra conversational or “chatty” response, you will need to explicitly ask for it.
- Multimodal Coherence: Textual content, photos, audio, or video ought to all be handled as equal-class inputs. Directions ought to reference particular modalities clearly to make sure the mannequin synthesizes throughout them relatively than analyzing them in isolation.
- Constraint Placement: Place behavioral constraints and function definitions within the System Instruction or on the very prime of the immediate to make sure they anchor the mannequin’s reasoning course of.
- Lengthy Context Construction: When working with giant contexts (books, codebases, lengthy movies), place your particular directions on the finish of the immediate (after the info context).
Gemini 3 Professional is our most superior mannequin for agentic coding. To assist builders get one of the best of its capabilities, we’ve labored with our analysis workforce to create a System Directions template for the mannequin that improved efficiency on a number of agentic benchmarks.
To start out constructing with these new options, take a look at the Gemini 3 documentation and skim the Developer Information for technical implementation particulars.
