Is Google Preparing Gemini 3.5 Pro+ Instead of Gemini 3.5 Pro? Fresh SDK Leak Points to “AGENTIC” Mode

Just when the tech community thought Google was finally ready to roll out its long-awaited flagship AI model, a fresh code update has thrown a fascinating curveball into the narrative.

For weeks, speculation reached a fever pitch around the impending release of Google’s next flagship model. Leaks covered by NokiaPowerUser tracked early momentum when Gemini 3.5 Pro was spotted on LM Arena, showing early builds competing directly against top rivals like Claude Fable 5. That momentum built quickly as reports suggested Gemini 3.5 Pro might drop today to challenge Claude Fable 5.

While expectations pointed to an August launch—with coverage noting Google might finally drop Gemini 3.5 Pro on August 12—the excitement hit a speed bump when Gemini 3.5 Pro was delayed again as last-minute glitches halted the expected drop.

Now, fresh commits pushed to Google’s official GenAI SDK repository hint at something even bigger: Gemini 3.5 Pro+.

First spotted by developer and AI commentator Choblin on X, the updated SDK logic reveals that while legacy models rely on traditional fixed-rate frame sampling for video processing, Gemini 3.5 Pro+ defaults to an entirely new AGENTIC mode driven by “model-driven dynamic navigation.”

What the Code Actually Tells Us

In standard multimodal video processing, AI models extract static image frames at uniform intervals—say, 1 frame per second. While simple, this fixed-rate method is inefficient. It wastes token capacity on repetitive footage and completely misses rapid, split-second action happening between samples.

The new SDK commit signals a fundamental departure from that paradigm:

// How Video Frame Extraction Is Shifting in Google's SDK
Older & Standard Models ---> Fixed-Rate Frame Extraction (1 fps, 2 fps, etc.)
Gemini 3.5 Pro+         ---> AGENTIC Mode (Model-driven dynamic navigation)

What Is “Model-Driven Dynamic Navigation”?

Instead of passively ingesting pre-sliced video clips, an agentic model acts like a human viewer with a video scrubbing bar.

• Intelligent Seeking: The model decides where and when to look based on context. It can skip past static, low-information scenes and home in on fast visual movements or key moments.

• Context Efficiency: By avoiding uniform frame dumps, context windows are preserved for high-value reasoning rather than redundant background pixels.

• Dynamic Interaction: When asked a complex spatial or temporal question (such as “At what point did the key slip out of the pocket?”), the model dynamically queries specific visual timestamps on demand.

Connecting the Leaks: How We Got Here

To understand why this SDK update is generating so much buzz, you have to look at how the timeline has unfolded over the past week:

1. The Arena Sightings: As detailed when Gemini 3.5 Pro was spotted on LM Arena, stealth builds surfaced quietly for benchmarking, sparking speculation that Google was preparing a surprise drop.

2. The Direct Rivalry: Industry reports highlighted how Gemini 3.5 Pro was positioned to challenge Claude Fable 5, particularly across multimodal understanding and coding benchmarks.

3. Target Launch Dates & Delays: While expectations initially set eyes on an August window—with reports suggesting Google might finally drop Gemini 3.5 Pro on August 12—the timeline hit a brief pause when Gemini 3.5 Pro was delayed again due to last-minute glitches.

Now, with references to 3.5 Pro+ appearing alongside native agentic video navigation, two clear theories emerge.

Two Theories: Pivot or Code Syntax?

Theory 1: An Upgraded Flagship Tier

If Google encountered unexpected engineering hurdles or chose to bundle its latest agentic navigation breakthroughs directly into the flagship release, Gemini 3.5 Pro+ could represent a revised, higher-tier release. Adding direct visual autonomy into the core model architecture would give Google a massive advantage in robotics, automated UI navigation, and real-time video analysis.

Theory 2: Wildcard Notation for Next-Gen Models

In software SDK development, appending a + sign often serves as shorthand or regex matching for “this model version and all subsequent iterations.” In this scenario, 3.5 Pro+ isn’t a separate product name, but a code flag ensuring that the core Gemini 3.5 Pro release—and any future spin-offs—automatically inherit the new AGENTIC video processing pipeline.

Why Agentic Multimodal AI Matters

The AI industry is rapidly moving away from simple text-in, text-out chatbots. The battleground centers on agentic execution—models that can see, reason, plan, and execute tasks across user interfaces and complex visual inputs autonomously.

If Google is baking “model-driven dynamic navigation” natively into its next Pro-tier model, it marks a shift from static media ingestion to active perception. Whether the final name on the product page reads Gemini 3.5 Pro or Gemini 3.5 Pro+, the underlying architecture is clearly preparing for a major leap forward.