AI Daily: Gemini 3 Flash Quietly Appears, Pro Version's Major Vision Breakthrough, and Antigravity Usage Limit Updates Explained

The AI circle has been incredibly lively these past few days. From Google DeepMind’s frequent moves, we are on the eve of a new wave of technological explosion. Whether it’s the mysterious model appearing in the arena or the significant leap in visual recognition technology, every piece of news touches the nerves of developers and tech enthusiasts. Ready to see what’s worth paying attention to today? Let’s take some time to talk about these ongoing changes.

1. Mystery Guest in the Arena: Gemini 3 Flash Suspected Exposed

If you have been paying attention to the LM Arena (Language Model Arena) recently, you may have noticed some unusual activity. Several models codenamed “skyhawk” and “seahawk” quietly appeared on the leaderboard, sparking heated discussions in the community.

Who exactly is this? Although the official announcement has not yet been made, all clues point to Google DeepMind’s next-generation lightweight model — Gemini 3 Flash. These two codenames are believed to be different checkpoints of the same series, one of which is likely the highly anticipated Flash version, and the other may be Flash Lite or even a more experimental version.

The significance behind this is actually quite interesting. With constant rumors about OpenAI’s GPT-5.2, Google’s choice to “warm up” a new model at this time is clearly preparation for the upcoming head-to-head confrontation. For developers, the Flash series has always represented the sweet spot of speed and cost. Whether this iteration can further improve reasoning capabilities while maintaining high efficiency is the part everyone is most looking forward to. After all, who doesn’t want an assistant that is both smart and quick to respond?

2. Gemini 3 Pro: A New Frontier for Visual AI

If Flash represents speed, then Gemini 3 Pro is undoubtedly the heavyweight show of strength. According to the latest technical details released by Google, this model has taken a huge step forward in “visual understanding”, and it can even be said that it is redefining how machines “see” the world.

Document Understanding: From “Recognition” to “Reconstruction”

In the past, our expectations for OCR (Optical Character Recognition) were at most to convert characters on an image into text. But how can real-world documents be so obedient? Scribbled handwritten notes, complex nested tables, and even yellowed manuscripts from a hundred years ago are all nightmares for traditional technology.

Gemini 3 Pro introduces a very powerful concept called “Derendering”. Simply put, it doesn’t just “read” the document, but understands the structural logic of the document. For example, when it sees a complex handwritten ledger from the 18th century, it can convert it into a perfectly structured table; seeing an image with mathematical formulas, it can directly output precise LaTeX code. This means that the model has the ability to reverse engineer visual information back to the original code (such as HTML or Markdown), which is definitely good news for digital archiving or automated processing.

Precise Reasoning and Spatial Awareness

In addition to understanding documents, this model has also learned to “think”. When processing long reports, it can cross-reference across dozens of pages of charts. Imagine throwing a 60-page census report at it and asking for the reason for the change in a certain data point over two years. It can find clues from the text narrative like a human analyst, then cross-reference the tables in the appendix, and finally give a comprehensive answer.

In terms of spatial understanding, Gemini 3 Pro demonstrates perception of the physical world. It can not only identify objects but also understand the “intent” of objects. This is widely applicable in the robotics field. For example, you can point to a messy table and ask a robot: “Help me figure out how to clean this up.” The model can generate a specific plan based on spatial coordinates. For AR/VR applications, this also means a more natural interactive experience.

Video Understanding: Understanding the “Why” Behind Actions

Video is the most complex type of data. Gemini 3 Pro has made two major upgrades in this regard:

1. High Frame Rate Capture: It can process video at a speed 10 times faster than standard (10 FPS). What is the use of this? Imagine analyzing a golf swing or a tennis serve; such millisecond-level action details can now be precisely captured by AI.
2. Causal Reasoning: This point is even more interesting. The model no longer just tells you “there is a person running in the picture”, but can understand “why they are running”. It introduces a mechanism similar to “thinking mode” to track complex causal relationships in videos. Even more, it can directly convert operational processes in long videos into executable code, which is simply a godsend for transforming instructional videos into practical applications.

For those interested in delving into the technical details, you can refer to the Detailed Introduction of Gemini 3 Pro.

3. Google Antigravity: Rights Adjustments for Paid and Free Users

As AI tools become more powerful, resource allocation has become a major issue. Google’s new development platform Antigravity recently announced new usage limit adjustments, which is a mixed bag for different groups.

A Boon for Paid Users, Growing Pains for Free Users

For Google AI Pro and Ultra subscribers, this is definitely good news. Officials have significantly increased the Rate Limits for such users and shortened the quota reset time to once every 5 hours. This means that professional developers do not have to worry about being forced to stop halfway through high-intensity projects, and productivity will be significantly guaranteed.

However, for users of the Free Plan, the rules have become much stricter. The limit has changed from the original short-term calculation to a weekly calculation. The official explanation is that this is to cope with the large amount of abuse and fraud encountered by Tier 1 (free user level), as these malicious traffic flows have affected the stability of paid content.

However, Google also mentioned that this is a “temporary” measure. They are developing more verification mechanisms and tiered plans to solve this problem. There is a small concept to establish here: in Antigravity, quota consumption is linked to the “workload completed by the Agent”. In other words, if you are just doing simple tasks, the consumption is actually not large; but if you let AI perform complex reasoning dramas, the quota will naturally be used up quickly.

4. Write Code to Win Prizes: Kaggle Vibe Code Challenge

Feel like you are good at writing Prompts? Or are you eager to try out the capabilities of Gemini 3 Pro? Now there is a chance for you to show off your skills. Google DeepMind is hosting a hackathon called Vibe Code with Gemini 3 Pro on Kaggle.

The core concept of this competition is “Vibe Coding” — using natural language prompts, combined with Gemini 3 Pro’s powerful reasoning and multimodal capabilities, to build applications in Google AI Studio that can solve real-world problems.

The most attractive part is undoubtedly the total prize pool of up to $500,000 (issued in the form of credits). The competition time is not long; it is a sprint that emphasizes creativity and implementation speed. If you have some crazy ideas, you might as well take this opportunity to implement them; maybe the next killer app will come from your hands.

5. API Policy Changes: The Exit of Gemini 2.5 and Future Prospects

Finally, we have to face some news on the reality front. The developer community discovered that the free quota for the Gemini 2.5 series API has changed drastically:

• Gemini 2.5 Pro’s Free Tier seems to have been removed.
• Gemini 2.5 Flash and Flash Lite’s daily requests (RPD) dropped precipitously from the originally generous 500 to 20.

** Note here: The original 500 times was what I found on the official website, but because the free tier has always fluctuated, I don’t know how much quota pro and flash originally gave, and the official has still not updated the latest news **

This is indeed catching people off guard, especially for students or individual developers who rely on free quotas for development and testing. The limit of 20 times can almost only be used to confirm “whether the connection is successful”.

What does this imply?

This is actually common product lifecycle management for cloud services. When a new generation of models (Gemini 3 series) is about to be fully rolled out, old resources are bound to be reallocated. Just like in the past, this is likely to free up server computing power for the upcoming Gemini 3.0 Flash.

We can boldly predict that after Gemini 3.0 Flash is officially released, Google will likely readjust the Free Tier strategy, and even offer a trial quota for Gemini 3.0 Pro to attract developers to migrate. Although it is a painful period now, in the long run, this is usually preparation for welcoming more powerful models.

Frequently Asked Questions (FAQ)

Q1: What is the difference between Skyhawk and Seahawk appearing on LM Arena? Although not officially confirmed, according to community speculation, these two should be different versions of the Gemini 3 Flash series. One might be the standard Flash version, and the other might be a Lite version with fewer parameters, or an experimental Checkpoint fine-tuned for specific tasks. Their appearance is mainly for “blind testing” before the official release to collect real human preference data.

Q2: What is “Derendering” and why is it important for document processing? Derendering is a core capability of Gemini 3 Pro. Traditional OCR can only recognize “there is a word on this picture”, but Derendering can understand “this word is a table header, followed by three lines of data”. It can reverse engineer visual images back into structured code (such as HTML or LaTeX). This means that AI is no longer just “reading words”, but can perfectly reproduce the layout structure and logic of complex documents, which is crucial for digitizing historical archives or processing financial statements.

Q3: Why did Google Antigravity change the limit for free users to “weekly calculation”? This is mainly to combat abuse. Google product managers pointed out that Tier 1 (free tier) has recently suffered a large amount of fraud and malicious traffic attacks, causing service instability and even affecting the rights of paid and legitimate users. Changing the limit to a weekly calculation and reducing the quota is to raise the threshold for malicious attacks and ensure system stability. This is a temporary measure and may be adjusted after the new verification mechanism goes online.

Q4: Now that Gemini 2.5’s free quota is so small, what should developers do? The current 20 RPD is indeed very tight. It is recommended that developers can:

1. Temporarily switch to using the free quota of the OpenRouter series (if still available).
2. If it is a commercial project, consider upgrading to a paid tier to ensure service stability.
3. Wait for the official release of Gemini 3 Flash, when new Free Tier plans will usually be released.
4. If testing, you can test directly on ai.dev.