Introduction

The AI race just got hot again out of nowhere. Gemini 3.2 Pro leaks, GPT 5.6 rumors, Mythos V2, and Codex mobile all appearing at the same time, and nobody saw it coming.

But here’s the thing though: Google hasn’t announced Gemini 3.2 officially. There’s no blog post, no API documentation, no changelog, nothing at all from their side.

But the community, developers, and pretty much everyone in the AI space already started hyping up what this model can do. Leaked test results are floating around, and honestly some of them look impressive.

On top of that, rumors about GPT 5.6 and Mythos are getting louder by the day, so this article will give you a full picture of where all 3 models stand right now, what to expect, and a rough timeline for when they might drop.

Some of this might be accurate and some might not, because after all none of us work inside Google, OpenAI, or Anthropic. What's covered overall for you:

I. Gemini 3.2 Pro Leaks: Is Google Back?

Leaked outputs show that Gemini 3.2 Pro significantly improves Scalable Vector Graphics generation. The model produces cleaner shapes and more stable layouts compared to the previous 2.5 version.

It can now generate interactive interfaces where users edit code and export files directly in the chat. However, the visual quality of generated user interfaces remains generic and lacks professional polish.

Key takeaways

• Leaked tests successfully generated complex vector graphics like a pelican riding a bicycle with clear proportions.

• The new SVG engine creates interactive customization tools rather than just static image files.

• Official announcements for Gemini 3.2 are widely expected to happen during Google I/O on May 19.

• Early outputs like a PS5 controller showed improved logic but still contained visible glitches in button alignment.

1. Where the improvement is real

The first leaked outputs of Gemini 3.2 Pro started showing up online recently, and they come from people who somehow got access to early test versions and shared the results publicly, not from any official Google demo.

This one focuses heavily on SVG generation, which stands for Scalable Vector Graphics. If you’re not familiar with SVG, think of it as drawing with code instead of pixels, so the image stays sharp at any size without getting blurry.

One of the first tests was generating a PS5 controller in SVG. The result looked okay, maybe a small step up from older Gemini models, but nothing that made people stop and stare.

The buttons didn’t line up properly, the overall shape felt a bit off, and the design had visible glitches.

My honest take is that the improvement is real but the jump feels small. That matters because the Gemini line has been quiet for months and expectations were sky high.

If you look more carefully at the leaked outputs, there’re real signs that Gemini 3.2 Pro improved its SVG generation. Another test showed a pelican riding a bicycle, and the result looked much more coherent than older Gemini models.

The bird had a clearer body shape, better proportions, and the bicycle itself was mostly complete instead of having broken parts or disconnected wheels.

The interesting part is that this wasn’t just a simple SVG image. The model also generated a full interface where you could customize the graphic, inspect the SVG code, export files, and edit different parts of the image.

This suggests Google may be focusing more on interactive SVG workflows instead of only basic image generation.

Overall, the SVG engine inside Gemini 3.2 Pro does look stronger than Gemini 3.1. The shapes are cleaner, the layout feels more stable, and the whole image looks more connected and natural.

Another leaked test suggests Gemini 3.2 Pro may be stronger at generating multiple SVG variations from a single prompt than people first expected.

In one example, the model created four different robot SVG designs, each with different colors, shapes, and small visual details instead of simply copying the same layout.

This suggests Gemini 3.2 Pro can generate more creative vector variations from the same idea, which could become a useful workflow for people working with icons, mascots, or vector assets.

3. Where it still falls short

When it comes to UI generation, the leaked outputs still look rough. The interfaces feel generic, the layouts look flat, and the components don’t have the polish you would expect from a flagship model in 2025.

To be fair, part of this could come from the prompts. If testers focused more on SVG quality than UI design, the results would naturally look weaker.

3. What to expect from Google I/O

Many people in the developer community expect an official Gemini 3.2 announcement around May 19, which is when Google I/O is scheduled. If that timeline holds, we'll have real benchmark data and official outputs to compare against these leaks within days.

Until then: SVG generation looks genuinely improved. UI generation remains the gap. Don't mistake better vector logic for high-quality visual design.

II. Gemini Flash Variants: Fanta, Sprite, and Cola

Key takeaways

• Gemini 3.2 Pro can generate multiple creative design variations from a single text prompt.

• Internal Flash variants use codenames like Fanta, Sprite, and Cola to test different optimization targets.

• The Fanta variant specifically showed highly competitive results against the more expensive Pro model.

• Pro remains the preferred choice for multi-step planning and tasks requiring deep context window usage.

1. Why Flash Variants Surprised Many Users

Separate from the Pro model, leaks have surfaced about the Gemini 3.2 Flash family, specifically, 3 internal variants with the codenames Fanta, Sprite, and Cola.

This kind of internal naming is standard practice. Labs typically train several configurations of the same base model, optimize each for different targets (speed, accuracy, specific task types), and then pick the strongest performer for public release.

What caught attention is how well one leaked variant performed. The Fanta variant showed benchmark results that are very close to Pro-level performance on several tasks.

If Google ships a Flash model that consistently performs near Pro quality at Flash prices, it changes the value calculation for a lot of developers who currently default to Pro for serious work.

2. What This Could Mean Before Google I/O

The expected post-I/O split:

• Pro handles complex reasoning, long-context tasks, and multi-step planning.

• Flash covers fast, practical work where speed matters more than depth. The question is how small the quality gap turns out to be.

If Google nails both releases they could reclaim a lot of attention. But if Pro doesn’t clearly outperform Flash in real world tests, the narrative will shift toward Flash being the real winner.

Until then, treat every leaked sample as a preview and not a verdict.

III. GPT-5.6 Already in Internal Testing

While attention is focused on Google I/O, OpenAI has apparently been running internal evaluations on GPT-5.6 for the past few days.

1. Ember Alpha And Beacon Alpha Model Names

→ There’re 2 internal codenames surfaced: Ember Alpha and Beacon Alpha, likely 2 variants being tested to see which gives better results across benchmarks.

OpenAI's typical pattern with internal candidates is to quietly deploy them on LM Arena or OpenRouter for blind public testing, or run A/B experiments inside ChatGPT where some users interact with the new model without knowing it. Based on past behavior, GPT-5.6 could surface publicly as soon as next month.

2. Why OpenAI Is Moving Faster: 2 Main Reasons

The speed here is remarkable. GPT 5.5 launched only a few weeks ago, so moving to GPT 5.6 testing this quickly means the gap between major releases is shrinking fast.

GPT 5.5 was a big success and many people said it outperformed Anthropic's Opus 4.7. That momentum gives OpenAI reason to push forward, especially with Google about to show new Gemini models at I/O.

If OpenAI waits too long, someone else grabs the spotlight.

The second reason is that AI itself now helps build AI. Models assist with code generation, data preparation, and evaluation during training runs. When your development tools improve every few months, development speed naturally increases.

This compression has a tradeoff: the massive capability jumps that defined transitions from GPT-3 to GPT-4 are less likely to repeat. Instead, updates focus on speed, efficiency, and workflow integration rather than dramatic new benchmarks.

IV. Mythos Preview V2 for Cybersecurity Models

Now let’s talk about something that doesn’t get enough mainstream attention. Mythos preview, which some people pronounce as Methos, just received a major upgrade.

The newer version is being tested by institutions including the AI Security Institute, and early benchmark results show a clear capability jump.

Here's how the cybersecurity AI rankings have moved recently:

• Claude Opus 4.6 led the evaluations

• Original Mythos Preview appeared but stayed behind Opus

• GPT-5.5 closed the gap, then OpenAI released GPT-5.5 Cyber, a security-tuned version that briefly took the top spot

• Mythos Preview V2 has now reclaimed the lead in current benchmark results

FYI, the most striking test involves a 32 step corporate network attack simulation, where the model needs to identify vulnerabilities, chain multiple exploits, move laterally through a network, and complete a full attack path.

A human security expert typically needs around 20 hours for this. Mythos Preview V2 completed it in roughly 6–10 attempts.

→ The same model that can defend a network can also be used to attack one. If Mythos Preview V2 can chain exploits in just a few attempts, that same ability could be misused on real systems.

That is why Anthropic launched Project Glasswing as a controlled rollout instead of a normal public release.

The company has already started giving governments, security agencies, and select organizations early access to Mythos Preview so they can prepare defensive responses before wider deployment.

V. Codex Mobile: Early Signs

The last piece is the smallest but worth noting for developers. Screenshots have started circulating that appear to show Codex running from a phone. Now in preview in the ChatGPT mobile app.

→ Official document.

What it looks like right now: a mobile CLI or remote coding workflow rather than a fully polished mobile app. That's still meaningful, because it means developers could potentially manage coding tasks, review outputs, and interact with Codex away from their desks.

Some earlier leaks also hinted at deeper integrations with tools like Notion, though nothing official has been confirmed yet. Rohan Burma from the Codex team was mentioned as a source, which adds credibility. This doesn’t look like random speculation.

There is also mention of a remote control feature coming soon, which would let you trigger and monitor Codex tasks from your phone while heavy processing runs on a server.

VI. Competition between OpenAI, Google, and Anthropic

The competition between OpenAI, Google, and Anthropic is becoming extremely aggressive. In several cases, major model launches happened within hours of each other as each lab tried to avoid losing momentum or attention.

When Mythos cybersecurity models started gaining traction, OpenAI quickly responded with GPT-5.5 Cyber. Now Google is preparing Gemini 3.2 while OpenAI is already testing GPT-5.6 only weeks after GPT-5.5 launched.

We, as normal users, are benefiting the most from this race:

• Models are becoming smarter and more practical

• Pricing is becoming more competitive

• Usage limits continue increasing

• Features are shipping much faster than before

• Tools like Codex and Claude Code are becoming more accessible to smaller teams and creators

The gap between people casually following AI and people actively adapting to these tools is growing very quickly.

If you want to stay ahead of these shifts, follow AI Fire for daily AI breakdowns, workflows, and practical trends.

Conclusion

The AI space is moving faster than most people can keep up with. Gemini 3.2 Pro leaks show SVG improvement but UI weakness. GPT 5.6 might already be on the way before GPT 5.5 settles in.

Mythos preview V2 is pushing cybersecurity AI to new levels. And Codex could soon work from your phone.

The real lesson is about the pace. When leaks appear before launches and every lab races to outdo the others, the field rewards people who stay informed and act on what matters.

Don’t treat every leak as gospel though. Wait for official releases, compare real outputs, and remember that sometimes the most meaningful improvements are the quiet ones that make your work faster and smoother without you noticing.

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

• 7 Practical Ways to Outperform 90% of Normal AI Users & 2x Your Productivity

• One Claude Prompt-Chain to Package Your Existing Skills Into a $1,000 Offer in 20 Days*

• The End of YouTube Tutorials? Gemini 3.1 Flash Changes Everything

• GPT-5.5's New Official Prompting Guide: Also Works for Claude & Gemini (Not Mega-Prompts)*

• ChatGPT Codex Replaced Hours of Complicated AI Video Editing. Just Prompt Instead*

  *indicates a premium content, if any

1