Google Gemini 3 Launches With Deep Think—But Only for Ultra Subs

Google rolled out Gemini 3 Pro to all Gemini app users on December 9, 2025, positioning it as their “most intelligent model yet” with state-of-the-art reasoning capabilities. The real headline, though, is what they’re holding back: Deep Think mode—the model’s most advanced reasoning feature—is locked behind the Google AI Ultra subscription. While OpenAI just shipped GPT-5.1 with adaptive reasoning to everyone, Google is creating a two-tier intelligence system where your wallet determines how deeply the AI thinks.

The timing isn’t coincidental. Just three weeks after OpenAI’s GPT-5.1 announcement, Google is racing to prove they can compete on reasoning tasks. But instead of matching OpenAI’s approach—making smarter models more accessible—Google is segmenting features by subscription tier. Gemini 3 Pro is free. Deep Think costs extra. The message is clear: Google wants recurring revenue, not just market share.

Gemini 3 arrives in a dramatically different landscape than Gemini 2. When Google launched Gemini 2 Flash in September 2025, they focused on speed and multimodal capabilities. The model was impressively fast but lagged on complex reasoning compared to GPT-4 and Claude 3.5 Sonnet.

Now, with reasoning models dominating developer conversations, Google needed an answer to OpenAI’s GPT-5 series and Anthropic’s rumored reasoning capabilities. Gemini 3 Pro is that answer—a model that “can reason across text, images, audio and video better than ever before” with improved coding abilities and what Google cryptically calls their “best vibe coding model yet.”

What does “vibe coding” mean? Your guess is as good as ours. Google’s tendency to invent marketing terms without clear definitions makes it hard to evaluate actual improvements. Based on the examples they provide—transcribing lecture notes, analyzing sports videos for form advice—it seems like they’re emphasizing multimodal understanding rather than pure coding performance.

The competitive pressure is obvious. Cursor, Cline, and other AI coding assistants have standardized on GPT-4 and Claude for serious development work. Google needs developers to believe Gemini 3 can compete, hence the vague but optimistic “best vibe coding model” claim.

For users in the standard Gemini app, Gemini 3 Pro brings several tangible improvements:

Better multimodal reasoning across text, images, audio, and video. Upload a video of yourself playing tennis, and Gemini 3 can analyze your form and suggest improvements. Record a lecture and get transcribed notes with context. The multimodal capabilities are genuinely impressive—Google’s been ahead of OpenAI here since Gemini 1.5.

More concise, better-formatted responses from the first interaction. Google claims outputs are now “more helpful” and better structured. In practice, this likely means they’ve fine-tuned away some of Gemini 2’s verbosity and improved markdown formatting. Small but welcome changes.

Access through the “Thinking” dropdown. To use Gemini 3 Pro, you select “Thinking” from the model menu. Confusingly, this is different from “Deep Think” (which is Ultra-only). Regular “Thinking” mode uses enhanced reasoning but not the iterative, multi-hypothesis approach of Deep Think.

What Google isn’t saying: how much “thinking” does standard Gemini 3 Pro actually do? OpenAI specifies token counts and reasoning effort levels. Google gives you a dropdown labeled “Thinking” with no visibility into what’s happening under the hood or how many resources are being consumed. For developers trying to optimize costs and latency, this opacity is frustrating.

“Higher limits” for paid subscribers. Google mentions AI Plus, Pro, and Ultra users get higher limits but never specifies what those limits are. Higher usage caps? Faster response times? Priority access during peak hours? The vagueness suggests these tiers might not be meaningfully different for most users.

Here’s where things get expensive. Gemini 3 Deep Think—the “most advanced reasoning mode”—is exclusive to Google AI Ultra subscribers. That’s $19.99/month, making it the most expensive consumer AI subscription on the market (ChatGPT Plus is $20/month but includes GPT-4 Turbo, DALL-E, and other features; Anthropic doesn’t have a consumer reasoning tier yet).

How Deep Think works: The model uses “iterative rounds of reasoning to explore multiple hypotheses simultaneously.” That’s Google-speak for what OpenAI calls “tree search” or what Anthropic might term “multi-path reasoning.” Instead of following one chain of thought, the model considers multiple approaches in parallel and synthesizes the best answer.

Performance claims: Google says Deep Think “excels at complex tasks including math, science and logic problems” and “pushes the boundaries of intelligence even further.” Note the careful language—they’re not claiming it beats GPT-5 or providing benchmark comparisons. Just that it’s better than standard Gemini 3 Pro.

The user experience trade-off: Deep Think takes “generally a few minutes” to respond. That’s slower than GPT-5 (which typically responds in 10-30 seconds for complex reasoning) but potentially more thorough. Whether the wait is worth it depends on whether Google’s parallel hypothesis exploration actually produces better results than GPT-5’s sequential reasoning.

Access method: Select “Deep Think” in the prompt bar AND “Thinking” in the model dropdown. This two-step process suggests Google is trying to make the feature feel premium and intentional, not something users accidentally trigger and then complain about wait times.

What’s conspicuously missing: any benchmark results. Where’s the SWE-bench score? MATH dataset performance? Comparison to GPT-5 on complex reasoning tasks? Google’s announcement is all qualitative claims without quantitative evidence.

Buried beneath the reasoning hype are two legitimately interesting features: visual layout and dynamic view. Both are experimental Labs features with limited rollout—”different subsets of users” will see them, which is Google’s way of saying “we’re A/B testing and not everyone gets access.”

Visual layout generates multimodal responses with photos and interactive modules instead of pure text. Ask for a 3-day Rome itinerary, and you get a visual, explorable plan with images and customization options across multiple turns. This is Google leveraging their Search index and multimodal capabilities to create richer outputs than text-only models can provide.

Dynamic view goes further—it generates custom user interfaces on the fly using “agentic coding capabilities.” Ask about the Van Gogh Gallery with life context for each piece, and Gemini designs and codes a unique interactive experience. Upload a photo of yourself for fashion advice, and you get a custom UI for exploring recommendations.

This is genuinely novel. OpenAI’s artifacts let Claude generate React components, but those are static templates. Google is claiming Gemini 3 can design bespoke UIs tailored to each specific prompt—essentially generative UI that adapts to content rather than using fixed layouts.

The catch? Both features are Labs experiments with limited availability. Google doesn’t specify rollout criteria—are they randomly selecting users? Testing by region? Limiting to paid tiers? The opacity suggests they’re not confident enough for full release, probably because the features work impressively 70% of the time and fail awkwardly 30% of the time.

Also unmentioned: performance implications. Generating custom UIs on the fly presumably requires significant compute. Does this impact latency? Consume more of your usage limits? Google’s silence suggests there are trade-offs they’d prefer not to highlight.

Google’s announcement is conspicuously light on concrete performance data. We get:

• Customer testimonials (carefully selected)
• Qualitative claims (“state-of-the-art reasoning”)
• Feature descriptions
• No benchmark scores

Compare this to OpenAI’s GPT-5.1 launch, which included:

• SWE-bench Verified scores (76.3%)
• Token efficiency graphs
• Speed comparisons with percentile breakdowns
• Specific customer metrics (Balyasny: “2-3x faster”)

Google’s approach suggests either (a) Gemini 3 doesn’t clearly win on standard benchmarks, or (b) they’re positioning this as a qualitative leap rather than quantitative improvement. Neither interpretation inspires confidence.

The lack of pricing transparency is equally frustrating. How much does Gemini 3 Pro cost through the API? What’s the token pricing for Deep Think mode? How do usage limits differ across tiers? Google’s consumer-focused announcement completely ignores developers who need this information to make decisions.

The headline mentions Gemini 3 rolling out to Search and Workspace, but Google’s announcement barely covers this. Reading between the lines: full integration is coming, but not immediately.

For Search: Gemini 3’s multimodal reasoning should improve complex queries that require synthesizing visual and textual information. Ask about comparing products with images, and Search could theoretically provide richer answers. But Google doesn’t specify when this happens or how it changes the search experience.

For Workspace: Gemini 3 could significantly improve Gmail drafting, Docs assistance, and Sheets data analysis. The multimodal capabilities mean you could upload a meeting recording to Docs and get structured notes. But again—no timeline, no specifics, no demo.

This feels like Google announcing aspirations rather than shipping features. “Rolling out” suggests gradual, controlled deployment, which translates to “most users won’t see this for months.”

Ultra subscribers are the obvious winners—Deep Think access for $19.99/month could be valuable for students tackling complex problems, researchers exploring hypotheses, or anyone doing serious analytical work. Whether it justifies the premium over free Gemini 3 Pro depends entirely on benchmark data Google hasn’t provided.

Multimodal power users will appreciate the improved reasoning across images, audio, and video. If you regularly work with diverse media types—analyzing video content, transcribing audio, processing images—Gemini 3’s multimodal strength is its clearest advantage over GPT-5.1.

Visual/dynamic view early access users get to experiment with genuinely innovative UI generation. If you’re selected for the Labs features, you’re essentially beta testing the future of generative interfaces. That’s exciting if it works, frustrating if you weren’t selected.

Who can skip this? Developers building on the API until Google provides pricing and benchmarks. Anyone satisfied with GPT-4 or Claude 3.5 Sonnet’s reasoning. Users who don’t need multimodal capabilities. And critically, anyone expecting Deep Think to be available without paying $20/month.

Available now: Gemini 3 Pro is accessible through the Gemini app’s “Thinking” dropdown for all users. Deep Think mode requires Google AI Ultra subscription ($19.99/month). Visual layout and dynamic view are rolling out to select users in Labs. According to Google’s official announcement, Search and Workspace integration is coming but without specified timelines.