OpenAI Codex Hits Mobile: The AI Programming Tool Ecosystem War Goes Full Scale

On May 15, 2025, the AI programming space saw a wave of major announcements: OpenAI integrated Codex into the ChatGPT mobile app, Microsoft began tightening Claude Code licensing, Tencent open-sourced an Agent Memory solution that dramatically reduces token consumption, NVIDIA unveiled its next-generation Agent compute platform, and a new company founded by a former Meta executive debuted at a $4.6 billion valuation. The ecosystem competition for AI programming tools has officially entered a white-hot phase.

OpenAI Codex Comes to Mobile, Windows Sandbox Paves the Way for Fully Autonomous Agents

OpenAI today officially embedded Codex into the ChatGPT mobile app, meaning developers can now generate code and debug on the go from their phones, no longer constrained to desktop environments. For developers who frequently need to review code during commutes or hotfix production bugs on short notice, this is an extremely practical update.

Codex is OpenAI's AI system fine-tuned specifically for code comprehension and generation tasks, built on the GPT series of large models. It first appeared in 2021 as the underlying engine powering GitHub Copilot, capable of understanding natural language instructions and converting them into executable code in multiple programming languages. From its initial API interface to now being directly embedded in a mobile app, Codex's product form has undergone a critical leap from developer tool to consumer-facing application.

Even more noteworthy is that OpenAI simultaneously released Windows sandbox technical documentation. A sandbox is a classic security isolation mechanism whose core idea is to create a restricted, virtualized execution environment within the operating system, preventing programs running inside it from accessing critical resources on the host system. Through privilege-elevated sandbox design, Codex is granted sufficient execution permissions within the Windows isolated environment—including installing dependencies, reading/writing files, and launching services—while ensuring these operations cannot escape to the host system, achieving an elegant balance between security and functionality. The intent behind this technical architecture is crystal clear—paving the way for fully autonomous PC Agents. When AI can not only write code but also safely execute it in an isolated environment, we're only one step away from a truly "fully autonomous AI programming assistant."

Microsoft Tightens Claude Code Licensing: Big Tech Ecosystem Walls Keep Growing Higher

While OpenAI charges ahead, competitors are having a tougher time. Microsoft has begun revoking some Claude Code licenses, directly impacting the experience of using Anthropic models in system components like Notepad++.

Claude Code is a command-line AI programming tool launched by Anthropic that allows developers to interact with the Claude model directly in terminal environments for code writing, debugging, and refactoring. Anthropic, one of OpenAI's primary competitors, was founded by former OpenAI core members, and its Claude series of models perform particularly well in code reasoning and long-context processing. Microsoft's move to tighten licensing is essentially the classic strategy of platform-level companies leveraging control over operating systems and development toolchains to shape the developer ecosystem—similar to the ecosystem lock-in logic seen historically in browser wars and mobile app stores.

Combined with the simultaneous preview launch of the GitHub Copilot App, Microsoft's intentions are unmistakable: to lock developers firmly onto its own platform through ecosystem binding. GitHub Copilot App, as Microsoft's proprietary AI programming assistant, deeply integrates with Azure OpenAI services, forming a complete closed loop from cloud compute to IDE plugins. The ecosystem walls between major players are growing ever higher, and developers genuinely need to consider multi-platform backup workflows. When your core development tools could be restricted at any moment due to commercial rivalry, risk management in technology selection becomes critically important.

Tencent Open-Sources TB Agent Memory: A Production-Ready Solution That Cuts Token Consumption by 61%

Tencent Cloud's database team today officially open-sourced the Tencent TB Agent Memory solution. This system has been deeply optimized specifically for long-task scenarios, achieving two impressive metrics through Mermaid task canvas and context co-loading technology: token consumption reduced by up to 61%, while task success rate actually increased by 51%.

To understand the value of this solution, you first need to understand the core challenge Agent Memory faces. When AI Agents (intelligent agents) execute complex multi-step tasks, they need to maintain large amounts of contextual information to remember previous operation steps, intermediate results, and task objectives. This information consumes the large model's context window in the form of tokens—tokens are the basic units by which large language models process text, with typically 1-2 tokens per English word and 1-3 tokens per Chinese character, and the cost of each API call is directly tied to token consumption. As task steps multiply, context keeps expanding, and token costs grow linearly or even super-linearly.

The Mermaid task canvas is a structured task representation method based on Mermaid diagram syntax that decomposes complex tasks into visualizable flowchart nodes, allowing Agents to store and retrieve task states in a more compact manner rather than retaining the entirety of verbose natural language conversation history. Context co-loading technology uses intelligent filtering and compression mechanisms to load only the contextual information truly needed for the current step into the model's input window, avoiding the inefficient traditional approach of cramming the entire conversation history into the prompt.

This solution supports mainstream frameworks including OpenAI and Claude, which is highly significant for development teams building complex Agent pipelines. One of the biggest pain points in deploying Agent applications today is persistently high token costs, especially in complex tasks requiring long-context memory, where a single task's API call fees can reach several dollars. If Tencent's open-source solution can truly deliver on its performance promises, it will dramatically reduce Agent application operating costs and accelerate commercialization.

NVIDIA Rubin Platform and Tiiny K2.6: Compute and Infrastructure for the Agent Era

NVIDIA has launched the all-new Vera Rubin platform—this is far more than just a hardware iteration. Named after American astronomer Vera Rubin, it inherits the design philosophy of the previous Blackwell architecture while further enhancing multi-chip interconnect capabilities. The platform specifically addresses the compute bottlenecks that Agent AI faces during scale-up, making it possible for thousands of Agents to collaborate simultaneously through optimized Pod architecture and interconnect technology.

Pod architecture refers to connecting multiple GPU accelerators via high-speed NVLink and NVSwitch into a compute cluster unit (Pod), with multiple Pods then connected via InfiniBand networking to form an elastically scalable ultra-large-scale compute pool. This architecture is particularly critical for Agent AI—unlike traditional single inference requests, large-scale Agent collaboration scenarios require extremely low-latency inter-chip communication and massive parallel inference capabilities, since each Agent may initiate independent inference requests simultaneously and needs to share intermediate states in real-time. For the compute foundation of the Agent era, NVIDIA clearly doesn't intend to let anyone steal this position.

On a related note, Tiiny K2.6's recent Agent mode has attracted widespread attention. It achieves "a production-grade database for everyone"—a website spun up from a single user sentence is backed by a complete independent database and backend hosting.

To handle million-scale database operations and maintenance demands, Tiiny has built out an ultra-low-cost Serverless database architecture. Serverless databases represent the cloud-native era's database delivery model, where users don't need to worry about server configuration, capacity planning, or operations management—database instances are automatically created on demand, elastically scale, are billed by usage, and can even scale down to zero cost when idle. This infrastructure capability has truly lowered the barrier to deploying AI applications to unprecedented levels—when database creation and operations can be automated by AI, independent developers' efficiency in building complete applications will see a qualitative leap. A database environment that previously took a backend engineer several days to configure can now be completed with a single natural language command.

RSI Scores $4.6 Billion Valuation: Having AI Research AI Itself

The final piece of news is a bombshell for the capital markets. Recursive Super Intelligence (RSI), the new company founded by former Meta executive Yuandong Tian after his departure, has just emerged from stealth mode with $650 million in funding at a valuation shooting straight to $4.6 billion.

Yuandong Tian is a prominent researcher in the AI field who spent years at Meta's (formerly Facebook) FAIR (Facebook AI Research) lab working on deep reinforcement learning and game-theoretic AI research, making significant contributions to Go AI, StarCraft AI, and other projects. He is among the few Chinese AI scientists who combine top-tier academic ability with industrial management experience.

Even more striking is the investor list—Jensen Huang (NVIDIA CEO) and Lisa Su (AMD CEO) both appear. The rare sight of both chip giant leaders jointly betting on the same startup reflects their strategic judgment about the explosion of AGI compute demand. RSI's goal is extremely ambitious: to train a system with the capabilities of 50,000 PhD-level researchers, enabling AI to automate research on AI itself.

This is essentially a bet on "Recursive Self-Improvement" as a path to AGI—the theory holds that if an AI system is sufficiently intelligent, it can understand and improve its own algorithms, architecture, and training methods, and the improved system then possesses even stronger self-improvement capabilities, forming a positive feedback loop that ultimately achieves exponential growth in intelligence. This concept can be traced back to mathematician I.J. Good's "intelligence explosion" hypothesis proposed in 1965. RSI's specific approach is to build an AI research automation platform capable of autonomously reading papers, designing experiments, validating hypotheses, and iterating improvements. If successful, it would fundamentally transform the paradigm of AI research. Of course, this is also a highly controversial direction, and its safety and controllability issues—including how to ensure the recursive improvement process doesn't diverge from human intent—are equally worth deep consideration.

Summary and Outlook

Looking at today's full news landscape, competition among AI programming tools has evolved from simple feature comparisons into an all-out war encompassing mobile experience, ecosystem binding, cost optimization, and compute infrastructure. OpenAI's Codex mobile + Windows sandbox combination, Tencent's open-source Agent Memory solution, NVIDIA's Agent compute platform, and RSI's recursive AI research vision together paint a picture of AI programming tools evolving at breakneck speed.

For developers, the most pragmatic advice is: embrace change but don't bet on a single platform, pay attention to open-source solutions to maintain flexibility, and closely watch the workflow transformation brought by Agent automation. The era of fully autonomous Agents may truly be closer than we think.