🤯 GPT 6 Could Be Taking Shape as Gemini 4 Gets Ready for the Next Big AI Fight

Introduction

🤯 Okay, the AI race is about to get messy again.

OpenAI is reportedly preparing GPT 6 with GPT Astra and Bell, while Google seems ready to bring Gemini 4 into the fight.

The funny part is that people are still getting used to today’s models, but the big labs are already lining up the next ones.

OpenAI may even have its own Jalapeño chip behind GPT 6, so this race is getting bigger than just “who has the smartest model.”

I’ll go through the biggest GPT 6 and Gemini 4 leaks, but keep one thing in mind: most of these details still haven’t been officially confirmed.

I. GPT 6 Could Be Built on Astra & Bell

OpenAI still hasn’t announced GPT 6, but the first pieces are starting to show up. Astra is the part we already know is real, while Bel makes the story behind it much more interesting.

Key points

• Astra may be an early step toward GPT 6.

• Bel is rumored to have 10T+ parameters.

• OpenAI hasn’t confirmed how Astra or Bel connect to GPT 6.

1. GPT Astra May Be The First Step Toward GPT 6

OpenAI officially called Astra “our next major model” and said an internal version of Astra produced new results on 10 long-standing math and computer science problems.

The community quickly connected Astra with GPT 6, especially after reports described Astra as one of OpenAI’s biggest pretrains since GPT-4.5.

OpenAI still hasn’t said that Astra is GPT 6. But when a new “major model” appears while everyone is waiting for GPT 6, the internet clearly isn’t going to wait for the official announcement.

2. Bel Could Show What OpenAI is Building After Astra

The next leak mentions Bel, a model described as Doug’s successor with more than 10 trillion total parameters.

According to reports tracking the Bel leak, OpenAI may continue reinforcement learning on Bel to build stronger systems, with Astra and future GPT models connected to the bigger roadmap.

Of course, 10 trillion parameters is enough for the timeline to start talking about AGI again. OpenAI still hasn’t confirmed Bel, its parameter count, or its exact role.

The bigger point is that OpenAI’s model pipeline may already be several steps ahead of what users can access today. GPT 6 hasn’t launched yet, but the Astra and Bel leaks suggest OpenAI may already be preparing what comes next.

II. OpenAI is Building Jalapeño To Support GPT 6

A model like GPT 6 may look very strong on benchmarks, but the harder question comes after that: how can OpenAI run such a large model for millions of users without making it too slow or too expensive?

Key points

• Jalapeño is OpenAI’s custom inference chip for future models.

• It could make interactions up to 4.1x faster and improve efficiency by up to 104.3x per watt.

• Better hardware could help OpenAI scale GPT 6 at lower cost.

That’s why Jalapeño matters. OpenAI is building this custom inference chip to support future models and reduce pressure on the compute behind them.

If Jalapeño works as expected, OpenAI could improve a few very practical things:

• Jalapeño could make interactions 2.1x to 4.1x faster.

• End-to-end latency could improve by 1.7x to 3.6x.

• Performance per watt could rise by up to 104.3x, helping OpenAI run models like GPT 6 more efficiently.

The problem is simple: bigger models can make compute costs rise very fast. GPT 6 will be harder to scale if OpenAI has to pay too much for every request.

Jalapeño suggests that OpenAI may already be preparing the infrastructure for its next generation of models. If this plan works, the GPT 6 race could also move deeper into the hardware behind the model.

III. Gemini 4 Could Become GPT 6’s Biggest Rival

OpenAI has GPT 6 in sight, while Google DeepMind is also reportedly preparing Gemini 4. Google has already confirmed that pre-training for the model has started.

If the leaks are right, Google isn’t just changing the model number. Gemini 4 could bring a much bigger upgrade.

Gemini 4 is rumored to have a 1.5 million-token context window, which could help it handle more documents, large codebases, and longer tasks in one session.

The reported upgrades also include:

• Better browser capabilities.

• Stronger terminal use.

• Improved coding workflows.

• Better tool use and multi-step agent tasks.

Some leaks also claim that Gemini 4 is doing very well in internal tests, even beating several current frontier models. Google still hasn’t released public benchmarks to confirm those claims.

If Gemini 4 really launches with these abilities, GPT 6 could have a serious rival from Google. The current model race hasn’t even cooled down yet, and the big labs already seem ready for the next round.

IV. What All These Mean for the Next AI Race

These leaks suggest that the next AI race won’t be only about which model gets the best benchmark score. As GPT 6 and Gemini 4 get bigger, compute and running costs may become just as important.

The key factors could include:

• Compute capacity.

• Inference speed.

• Reliability.

• Infrastructure.

• Cost-performance.

DeepSeek V4 Flash is a good example.

According to DeepSeek’s official benchmark and pricing pages, the model scored 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE, while its API price was once as low as $0.14 per 1M input tokens and $0.28 per 1M output tokens.

A model doesn’t need to win every benchmark if it can still deliver strong performance at a much lower price.

Anthropic is also spending heavily on compute. Reuters reported that Anthropic has a $45B deal with Nscale for more AI computing capacity, while the company is also paying for extra infrastructure elsewhere.

If OpenAI moves fast with GPT 6 and Google comes back strongly with Gemini 4, Anthropic may have to keep up with both model quality and the compute needed to run these systems.

The next AI race may look less like a simple benchmark chart and more like a fight over who can run powerful models at a price people can actually afford.

V. What’s Unknown about GPT 6 & Gemini 4

This is where we need to bring the hype back down a little. The leaks sound exciting, but many important details still haven’t been confirmed.

The biggest questions are still open:

• Is GPT Astra really connected directly to GPT 6, or is it a separate model in OpenAI’s roadmap?

• Does Bell really have more than 10 trillion parameters, as the leaks claim?

• Will Jalapeño only support part of OpenAI’s inference system, or become important infrastructure for GPT 6?

• Will Gemini 4 really have a 1.5M context window with stronger browser, terminal, and tool use?

• Will the internal benchmark results still look strong when the models are tested publicly?

We already have many pieces of the story, but there still isn’t enough proof to build the full picture.

The internet may be a few steps ahead of the official announcements, while the final answers still depend on what OpenAI and Google actually release.

Conclusion

The next AI race already looks crowded. GPT 6 may be taking shape behind Astra and Bel, OpenAI is building Jalapeño to handle the compute side, and Google is preparing Gemini 4 with stronger tool use and a much larger context window.

The funny part is that none of the biggest pieces are fully confirmed yet, but the industry is already acting like the next round has started.

For now, the real question isn’t only which model will be smarter. It’s which company can make its model fast, reliable, and cheap enough for people to actually use at scale.

If even half of these leaks are true, GPT 6 vs Gemini 4 could become one of the biggest AI battles of the year.

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:

• ChatGPT Just Got a New Superpower? (Computer History & More Updates)

• Grok Bot Just Killed 1,000 Startups? An Insanely Easy AI Agent Anyone Can Use 24/7

• Want to Sell AI Workflows? Start With These 5 High-Demand Business Automation in 2026*

• Easily Build Your Own AI Agent Team with Claude Code & Codex in One Workspace*

• NEW ChatGPT Work is the Claude Cowork Killer? Step-by-Step Analysis & Full Review*

  *indicates a premium content, if any