Gemini 3 Flash beats Pro... Are we Cooked?

Gemini 3 Flash: Detailed Overview & Analysis

Key Concepts:

• Gemini 3 Flash: Google DeepMind’s new flash model offering pro-level performance at lower cost and latency.
• Dynamic Thinking: A system where the model dynamically allocates reasoning resources based on query complexity.
• Sweet Bench Verified: A coding benchmark used to evaluate model performance.
• Context Window: The amount of text a model can consider at once (1 million tokens for Gemini 3 Flash).
• Multimodal Inputs: The ability to process different types of data like text and images.
• API (Application Programming Interface): A set of rules and specifications that software programs can follow to communicate with each other.
• Function Calling/Tools: Allowing the model to utilize external tools like Google Search.
• Reasoning Effort/Levels: Controls the depth of reasoning the model applies (Minimal, Low, Medium, High).

---

1. Introduction & Core Capabilities

Google DeepMind has released Gemini 3 Flash, a significant update aiming to deliver “pro-level performance” with the speed and cost-effectiveness of a flash model. Historically, flash models prioritized low latency and cost but lacked the intelligence of more powerful models. Gemini 3 Flash bridges this gap, particularly excelling in coding and reasoning tasks. Early access testing indicates this is one of the most impressive flash model releases to date. The model is currently available in preview via AI Studio, Vortex, Google Anti-gravity, and the Gemini CLI, using the model string “Gemini 3 flash preview.” Multimodal inputs are supported with adjustable media resolution (low, medium, high) impacting token count and accuracy.

2. Dynamic Thinking & Reasoning Levels

Gemini 3 Flash introduces “dynamic thinking,” a novel approach to reasoning. Unlike fixed reasoning levels, the model assesses each query and allocates a “thinking budget” accordingly. Four reasoning levels are available – Minimal, Low, Medium, and High – but the model itself determines the actual reasoning effort. Setting the level to “Minimal” doesn’t guarantee no reasoning, but aims to minimize it. This is crucial for API integration, requiring developers to account for potential reasoning tokens even with minimal settings. The model retains the 1 million token context window of previous generations.

3. Benchmarking & Performance Analysis

Initial benchmarks demonstrate Gemini 3 Flash’s strong performance. Notably, it surpasses Gemini 3 Pro on the Sweet Bench verified coding benchmark, a consistent trend suggesting it’s a specialized coding model. Compared to Gemini 2.5 Pro (state-of-the-art in March), Gemini 3 Flash shows significant improvements.

• Sweet Bench Verified: Gemini 3 Flash outperforms Gemini 3 Pro and significantly exceeds Gemini 2.5 Pro.
• GPQA Diamond: Performance is very close to Gemini 3 Pro and surpasses Gemini 2.5 Pro, testing scientific and reasoning capabilities.
• Humanity’s Last Exam: Performance is comparable to Gemini 3 Pro, exceeding models like Sonnet Pro 4.5 and GPD 5.1.
• Coding Tasks (vs. Gemini 3 Pro): For well-defined tasks, Gemini 3 Flash can effectively replace Gemini 3 Pro for planning.

4. Pricing & Cost Considerations

While offering superior performance, Gemini 3 Flash comes with a price increase compared to Gemini 2.5 Flash. The cost has risen from $0.30 to $0.50 per million input tokens, with a smaller increase for output tokens. However, the model still provides performance very close to Gemini 3 Pro at a significantly lower cost, and remains competitive with other providers. Google’s control over the entire machine learning stack (training and inference) enables this cost efficiency.

5. Comparative Examples & Demonstrations

The video showcases direct comparisons between Gemini 3 Flash, Gemini 2.5 Pro, and Gemini 3 Pro.

• Spatial Awareness & Coding: A prompt testing both coding and spatial reasoning revealed Gemini 2.5 Pro failed to recognize a TV in a scene, while Gemini 3 Flash accurately depicted Tom and Jerry on the screen. Gemini 3 Pro produced a more aesthetically pleasing result but took significantly longer (22s for Flash vs. 37s for 2.5 Pro vs. 48s for 3 Pro).
• Website Design: Gemini 3 Flash, with effective prompting (using URL context to analyze an existing website and requesting a “billion-dollar design company” reimagining), generated a well-designed website comparable to Gemini 3 Pro’s output. Simple prompts yield typical AI-generated results, highlighting the importance of strategic prompting.
• Animation & Video Sequence Creation: Gemini 3 Flash created a video sequence with animations in 40 seconds, while Gemini 3 Pro took 90 seconds. While Gemini 3 Pro’s output was more detailed, Gemini 3 Flash achieved good results with appropriate prompting.
• SVG Creation: Gemini 3 Flash successfully generated an animated SVG of Spongebob with controllable hand movements, all within a single file.

6. Advanced Capabilities & Reasoning Tests

Gemini 3 Flash demonstrates impressive spatial understanding and reasoning, mirroring Gemini 3 Pro.

• Sentence Arrangement: Like Gemini 3 Pro, it consistently arranged people into a sentence ("hello world, I am Gemini"), a task few models can perform reliably.
• Simulated MacOS: Gemini 3 Flash created a functional, aesthetically pleasing simulated MacOS environment in a single HTML file, including a working calculator and text editor. Some features (folder switching) require further prompting refinement.
• Modified River Crossing Puzzle: While struggling initially, Gemini 3 Flash demonstrated a unique approach, recognizing the emphasis on the goat’s well-being and proposing a complex solution. It came closer to the correct answer than many other models.
• Modified Trolley Problem: Correctly identified the ethical implications of the scenario, recognizing the people on the track were already deceased and therefore not requiring intervention.

7. API Integration & Usage

Gemini 3 Flash can be used as a drop-in replacement for other models via the API, using the model string “Gemini 3 flash preview.” The API supports tools/function calling (e.g., Google Search) and the four reasoning levels. A detailed Google notebook provides guidance on reasoning models and API usage. Crucially, the model dynamically manages its thinking budget, even when set to “Minimal.”

8. Recommended Usage Pattern & Conclusion

The presenter recommends using Gemini 3 Flash as a “workhorse” model in conjunction with Gemini 3 Pro. Utilize Gemini 3 Pro for planning, design, and orchestration, then leverage Gemini 3 Flash for efficient implementation of well-defined features. This approach balances speed, cost, and capability. Gemini 3 Flash is described as being comparable to Haiku 4.5 in relation to Opus 4.5. The key takeaway is that Gemini 3 Flash offers a compelling combination of performance, speed, and cost-effectiveness, making it a valuable addition to the AI ecosystem. Continued updates and refinements are expected, as seen with previous Gemini models.

Notable Quote:

“You get pro-level performance at the cost and latency of a flash model which is pretty impressive.” – Presenter, describing the core benefit of Gemini 3 Flash.