OpenAI and Google DeepMind are preparing their next generation of frontier artificial intelligence models as competition at the top of the AI industry continues to intensify. OpenAI has publicly identified Astra as its next major model, while reports circulating around the project claim that it could eventually appear as “GPT-6 Astra” and use an architecture containing more than 10 trillion parameters.
Those specifications remain unconfirmed, however. OpenAI has not publicly disclosed Astra’s parameter count, confirmed the GPT-6 branding or detailed a so-called “Bell” architecture. Google, meanwhile, has acknowledged work on Gemini 4, but several reported features and release details also remain unofficial.
The emerging picture nevertheless shows where the frontier AI race is heading: larger and more capable models, much longer tasks, deeper integration with tools and browsers, and increasingly specialized computing infrastructure.
Summary
Reports surrounding OpenAI’s upcoming Astra model claim it could use a new “Bell” architecture containing more than 10 trillion parameters, although OpenAI has not confirmed either specification. The company has officially described Astra as its next major model and has already demonstrated an internal version on advanced mathematics and theoretical computer science problems.
Google is also developing Gemini 4 as its next-generation AI platform. Reports point to improved browser capabilities, deeper tool integration and a larger context window, but these individual specifications have not yet been officially confirmed.
At the infrastructure level, OpenAI has already moved beyond rumor with Jalapeño, its first custom inference chip. The processor is designed to improve throughput and energy efficiency as the cost of operating increasingly powerful AI systems becomes a strategic issue.
OpenAI’s Astra could represent a major architectural shift
Reports cited by Universe of AI describe Astra as being built around a new model architecture known internally as “Bell,” allegedly succeeding an earlier system called “Doug.” The reports claim the architecture could contain more than 10 trillion parameters, potentially making it one of the largest AI systems ever developed.
For now, those figures should be treated as reported rather than confirmed specifications. OpenAI has publicly referred to an internal version of Astra as its “next major model,” but the company has not published a model card detailing its parameter count, architecture, pricing or eventual product branding.
The distinction matters because headline parameter counts no longer provide a complete picture of model performance. Modern frontier systems can use techniques such as expert routing, tool use, retrieval and extended reasoning, meaning the total number of parameters may differ substantially from the amount of computation used for each individual request.
Larger context and longer-running AI tasks
The reports also point to an expanded context window and a greater ability to process large datasets while making decisions across longer workflows. Such capabilities could be particularly important in healthcare, financial analysis, scientific research and logistics, where AI systems increasingly need to combine information from multiple sources rather than simply answer isolated prompts.
OpenAI’s public demonstrations of Astra already suggest a greater emphasis on complex reasoning. An internal version of the model has been used on advanced mathematics and theoretical computer science problems, indicating that long-horizon research tasks are becoming an important part of the company’s frontier-model strategy.
TL;DR Key Takeaways
- OpenAI has confirmed Astra as its next major model, but reports that it will be called “GPT-6 Astra,” use a “Bell” architecture or contain more than 10 trillion parameters remain unconfirmed.
- Google DeepMind is developing Gemini 4, with reports pointing to stronger browser and tool integration and a substantially expanded context window.
- OpenAI’s custom Jalapeño inference chip is designed to improve AI performance and energy efficiency, highlighting the growing importance of specialized hardware.
- Global AI competition continues to intensify as Chinese developers including DeepSeek and GLM challenge established U.S. laboratories.
- The next phase of the AI race is increasingly about the combination of models, tools, infrastructure and energy efficiency rather than raw model size alone.
Gemini 4 becomes Google DeepMind’s next major frontier model
Google has publicly acknowledged that Gemini 4 is under development as part of its next generation of AI systems. The project arrives as competition from OpenAI and Anthropic pushes leading laboratories to improve reasoning, coding, agentic workflows and the ability to work across extremely large amounts of information.
Reports surrounding Gemini 4 suggest that Google is focusing heavily on tool integration and more autonomous workflows. Among the capabilities being discussed are:
- Improved browser capabilities, enabling more seamless interaction with websites and online information.
- Terminal integration designed to support more advanced tool usage and automation.
- A reported context window of up to 1.5 million tokens for complex workflows and deeper analysis.
These individual specifications have not all been publicly confirmed by Google, so they should be viewed as indications of the development direction rather than final product specifications. Google has, however, described the Gemini 4 generation as part of its continuing push at the AI frontier.
Why browser and terminal integration matter
The importance of deeper browser and terminal access goes beyond convenience. Frontier AI systems are increasingly evolving from conversational assistants into agents capable of gathering information, manipulating files, using software and completing multi-stage tasks.
A model with reliable access to browsers, terminals and other external tools could therefore perform workflows that previously required constant human intervention. For businesses and developers, this could make AI more useful in areas such as research, software engineering, financial analysis and business-process automation.
A larger context window would complement those capabilities by allowing the system to keep considerably more information available during a task. That could become especially important as AI agents move from minutes-long interactions toward workflows lasting hours or even days.
OpenAI’s Jalapeño chip adds a hardware dimension to the AI race
OpenAI is also investing directly in the hardware required to run frontier AI models. Its first custom inference processor, Jalapeño, was developed with Broadcom and is designed specifically for serving trained AI systems rather than training them.
The strategic objective is straightforward: deliver more AI inference while consuming less electricity and reducing latency. Recent benchmark results presented by OpenAI suggest that Jalapeño can deliver significantly more work per unit of power than some existing accelerator systems under selected inference workloads.
That matters because the AI race is increasingly constrained by electricity, data-center capacity and semiconductor availability. Improving intelligence per watt can lower operating costs while allowing companies to serve significantly larger numbers of users from the same infrastructure.
Custom chips become a strategic advantage
The development of Jalapeño reflects a broader shift across the technology industry. Google has spent years developing its own TPUs, while Amazon, Microsoft and other major cloud operators are also investing in specialized AI processors.
For OpenAI, custom silicon provides another way to optimize its infrastructure around the specific demands of its models. The company is not abandoning external chip suppliers, but developing its own processors could reduce some dependence on general-purpose AI accelerators and provide greater control over cost and performance.
Hardware innovation may therefore become almost as important as model architecture in determining which AI companies can scale economically.
Global competition expands beyond OpenAI and Google
The frontier AI race is no longer confined to a handful of U.S. laboratories. Chinese developers including DeepSeek and GLM continue to improve their models, increasing competitive pressure on OpenAI, Google DeepMind and Anthropic.
Access to computing resources remains one of the major constraints. Advanced chips, data-center infrastructure, electricity supply and the ability to operate models efficiently at scale are becoming critical strategic resources.
This dynamic could increasingly favor organizations that can optimize the entire AI stack rather than simply train the largest model. Model design, custom hardware, inference efficiency, data-center capacity and agent infrastructure are becoming interconnected parts of the same competition.
The race is shifting beyond parameter counts
The reported scale of Astra attracts attention because a model containing more than 10 trillion parameters would represent an extraordinary engineering project. Yet raw parameter count is becoming a less useful measure of practical AI capability.
The more important question is how effectively those resources translate into reasoning quality, reliability, tool use and usable performance. An enormous model that is prohibitively expensive to operate may ultimately provide less commercial value than a smaller system that can complete complex tasks faster and at lower cost.
This is why OpenAI’s work on Astra and Jalapeño should be viewed together. Frontier-model development and infrastructure optimization are increasingly two sides of the same strategy.
Implications for the future of AI
The next generation of AI systems could have major consequences for industries that depend on large amounts of information and repeated decision-making. Healthcare, finance, logistics, scientific research and software development are among the areas most likely to benefit from models capable of maintaining longer context and operating external tools.
At the same time, increasingly autonomous systems introduce new challenges around reliability, security and oversight. The ability to browse the web, execute commands or operate software makes a model more useful, but it also raises the stakes when that model makes an incorrect decision.
The pursuit of increasingly general AI capability therefore involves more than benchmark improvements. Efficiency, controllability and safe deployment will become increasingly important as models gain more autonomy.
The path forward
Astra and Gemini 4 represent the next stage of a competition that is rapidly expanding from model intelligence into complete AI platforms. Models, agents, browsers, coding tools, custom chips and data-center infrastructure are increasingly being developed as parts of integrated systems rather than isolated products.
Many of the most striking specifications currently circulating around these models remain unconfirmed, especially Astra’s reported 10-trillion-parameter scale and several of Gemini 4’s proposed features. Until OpenAI and Google publish full technical details, those claims should therefore be treated cautiously.
What is already clear is that the competitive frontier has moved beyond simply building a better chatbot. The leading AI laboratories are now competing to create systems capable of reasoning, using tools and operating across increasingly complex real-world workflows, while simultaneously reducing the enormous infrastructure costs required to run them.
