Google's most powerful Gemini model remains stuck in testing while its cheaper, faster alternatives keep shipping new versions. As of September 2026, Gemini 3.5 Pro has not been released to the public, despite being announced as coming "the following month" at Google I/O in May. Meanwhile, Google debuted Gemini 3.8 Flash just three weeks after releasing Gemini 3.7 Flash, signaling that the company's fastest progress is happening in the budget tier, not the premium one.

Why Is Google's Top-Tier Model Delayed?

The gap between Gemini's two main tiers has become impossible to ignore. Google positions Gemini Pro as its highest-capability offering for complex reasoning and coding tasks, while Flash prioritizes lower cost and faster response times for production applications. The problem: Pro is currently stalled. Google DeepMind's Logan Kilpatrick acknowledged that the company is still testing Gemini 3.5 Pro with partners following internal delays, and the current flagship Pro model remains Gemini 3.1 Pro from February 2026.

This creates an unusual situation in AI development. Typically, companies race to release their most powerful models first, then optimize cheaper versions later. Google appears to be doing the opposite, at least for now. The practical takeaway is clear: Google's fastest progress is in the cheaper, faster tier, while the top-capability tier is currently stalled.

What's Actually Shipping in Google's AI Lineup?

While Gemini Pro waits in the wings, Google has been remarkably productive elsewhere. The company's AI model catalog now spans text, images, video, audio, music, interactive 3D worlds, and physical robots. Rather than treating these as separate products, Google frames them as a general-purpose engine with specialized attachments. Gemini is the engine; Omni, Nano Banana, and Audio are Gemini-branded attachments for particular output types.

The most architecturally significant launch of 2026 is Omni, Google DeepMind's first natively multimodal generative media model. Unlike previous systems that ran separate models for video (Veo), images (Imagen), and audio, Omni combines them into one model that reasons across all three modalities at once. This structural change means more coherent edits and fewer artifacts from handing work between systems. The main user-facing feature is conversational editing: each follow-up prompt edits the same scene instead of generating a new one, so characters, lighting, and continuity carry across turns.

How to Navigate Google's Expanding AI Model Ecosystem

• For Text and Reasoning Tasks: Gemini Flash models offer frontier performance on agents, coding, and complex long-horizon tasks at lower cost, while Gemini 3.1 Pro remains the highest-capability option for maximum reasoning power, though it has not been updated since February 2026.
• For Image Generation: Nano Banana 2 Lite generates images in about four seconds at $0.034 each, making it cost-effective for high-volume production, while Nano Banana Pro remains the choice for high-fidelity work requiring maximum factual accuracy.
• For Video and Multimodal Work: Gemini Omni Flash is now live as Google's latest video model and will replace Veo in the Gemini app, offering conversational editing across video, images, and audio in a single model.
• For Audio and Speech: Gemini 3.8 Live and 3.8 Live Extended Thinking represent a step change in native speech-to-speech models that can carry out tasks while keeping dialogue going, with the Extended Thinking variant ranking first on Artificial Analysis' Speech-to-Speech leaderboard.
• For Music Generation: Lyria 3 Pro generates complete compositions up to three minutes long and understands song structure, while Lyria 3.5 launched in July with improvements in musicality, lyrics, and vocal quality.

Google's strategic trend in 2026 is that specialized attachments are being pulled back into the engine. Omni is displacing its predecessor: Google describes Omni as its latest video model and says it will replace Veo in the Gemini app. Developers should note that the gemini-omni-flash-preview endpoint will be deprecated on September 30, 2026.

The audio family has also expanded significantly. Gemini 3.5 Transcribe offers accurate, low-latency speech-to-text across more than 85 languages, with speaker diarization, word-level timestamps, and custom vocabulary of up to 1,000 terms. Gemini 3.5 Live Translate provides speech-to-speech translation across more than 70 languages. Text-to-speech was upgraded recently: Google introduced Gemini 3.8 Flash TTS, built for detailed creative direction and character design, which can create new voices from scratch using natural-language prompts. All audio from Gemini Audio models carries an imperceptible SynthID watermark so AI-generated speech stays detectable.

What About Google's Specialist Models?

Beyond the Gemini family, Google maintains specialist systems for specific tasks. Genie is in a different category from video generators. While video generators produce a clip you watch, Genie produces an environment you move through, generated in real time in response to your actions. A world model simulates how an environment behaves, predicting how it evolves and how actions affect it. The difference is like the difference between a film and a video game: one is fixed playback, the other responds to input.

Genie 3 was released in August 2025 with higher-resolution worlds and several minutes of visual consistency. On January 29, 2026, DeepMind opened access to AI Ultra subscribers through Project Genie. At Google I/O 2026, Project Genie added a Street View integration that generates navigable simulations of real-world locations from Street View data. The model's biggest value may be industrial rather than recreational: Waymo built a variant called the Waymo World Model to simulate edge cases for its robotaxis.

Lyria, Google's music model, released Lyria 3 Pro in 2026, which generates complete compositions up to three minutes long and understands song structure such as intros, verses, choruses, and bridges. Standard Lyria 3 produces tracks up to 30 seconds for prototyping and short-form use. On the copyright question central to AI music, Google says Lyria 3 Pro outputs are watermarked and designed not to imitate existing artists.

For developers who want to run models on their own hardware, Gemma is Google's open-weights version of its AI models. This allows researchers and companies to avoid cloud dependencies and maintain more control over their AI infrastructure.

Looking ahead, Google DeepMind has begun its most ambitious pre-training run yet for Gemini 4, suggesting that while Gemini 3.5 Pro remains in testing, the company is already preparing its next major leap. The delay in Gemini 3.5 Pro's release underscores a broader truth about AI development in 2026: the race to build the cheapest, fastest models is moving faster than the race to build the most capable ones.