A New Starting Point and the Real Journey for the AI Industry

Recently, whether in Silicon Valley developer mailing lists or domestic tech communities, discussions have been dominated by two major pieces of news.

One is OpenAI’s official confirmation that the next-generation large model, GPT-6, codenamed Spud, is set for a global release. The other is that the company, just before the release, is shutting down Sora, the video generation product that took the internet by storm a year ago, and has even set a timeline for the complete deprecation of its API.

Many AI developers have been adjusting the pace of their projects over the past two weeks, waiting to see GPT-6’s release documentation. Some are also discussing whether OpenAI is moving too hastily—after all, when Sora was launched, it almost redefined the standard for AI video generation. How could it be shut down so abruptly?

In fact, rather than focusing on how many parameters GPT-6 can stack or how much its reasoning ability has improved, it’s worth examining the series of choices OpenAI made before this model’s release, as well as the subtle changes quietly happening across the entire AI industry ahead of this launch.

01

The Essence of GPT-6 and OpenAI’s Strategic Shift

In OpenAI’s recent official announcement, there wasn’t much information about GPT-6—only confirmation that pre-training work has been fully completed at the Stargate data center in Texas, and the model is now in the final stages of safety alignment and API debugging.

More details emerged from sporadic statements by core management.

OpenAI President Greg Brockman confirmed GPT-6’s existence in a recent podcast interview. His remarks avoided the usual focus on parameter counts, leaving only one key judgment: this is not an incremental improvement. The way they think about model development has undergone a major transformation.

Sam Altman, in an internal letter to all employees, defined this model as “a very powerful model that can truly accelerate economic development.”

From leaked test information within the industry, GPT-6’s core upgrades indeed break away from the inertia (inertial) path of previous generations. It adopts a brand-new “Symphony” architecture, achieving native unified processing of text, images, audio, and video for the first time—rather than the common industry approach of stitching together multimodal modules.

This means users no longer need to jump through plugins. They can directly generate front-end code from hand-drawn sketches, upload a video to dissect motion details and generate corresponding scripts, or even complete the entire workflow from creative conception to finished video using voice commands.

For developers, the most notable improvements are in coding, reasoning, and AI agent tasks. GPT-6 delivers over 40% performance gains compared to GPT-5.4, with its context window expanding from 1 million tokens to 2 million, allowing it to process approximately 1.5 million words of text in a single session—equivalent to two novellas.

Public test data shows its mathematical reasoning accuracy reaches 92.5%, code generation pass rate hits 96.8%, and in 44 professional test categories, 83% of tasks perform at or near human expert levels.

What’s even more noteworthy are the preparatory moves OpenAI made for this release, especially the complete shutdown of Sora. I remember when Sora first emerged, the industry called it a “game-changer.” Many video creators were amazed by its capabilities.

But when it came to turning text-to-video generation into viable commercial projects, very few succeeded. Either the copyright of generated content was unclear, or costs were too high—far less controllable than human-made productions.

This was the core reason behind Sora’s shutdown. According to Forbes estimates, Sora’s annual operating costs exceeded $5 billion, while its total in-app revenue since launch was only about $2.1 million—nowhere near covering its massive computational and operational expenses.

Unlike text generation, AI video generation consumes exponentially more computing power. Generating a basic 10-second video costs about $1.30, while complex scenes can reach $33.

To control losses, OpenAI had to repeatedly cut users’ free generation quotas, from 30 daily downloads initially down to just 6, further accelerating user churn. Appfigures data shows Sora’s 30-day retention rate was only 1%, and 60-day retention approached zero—users tried it once but had little incentive to keep using it.

Sora also faced crushing pressure from copyright and compliance issues. Early on, it quickly gained attention by generating videos featuring Disney IPs and celebrity likenesses, but this triggered numerous copyright lawsuits and industry backlash.

OpenAI was forced to tighten content generation rules from “default usable” to “requires explicit authorization,” directly undermining the product’s core appeal. Its $1 billion partnership with Disney also terminated. The proliferation of deepfake content further subjected OpenAI to regulatory pressure, with the U.S. nonprofit organization Public Citizen specifically calling for Sora’s removal.

Of course, some also speculate that OpenAI is rushing toward an IPO. The company completed a new $122 billion private funding round in late February 2026, reaching an $852 billion valuation, with plans to go public in Q4 2026. For Wall Street investors, a money-losing video generation project with no clear path to profitability is far less attractive than an enterprise productivity model that can generate stable cash flow.

Goldman Sachs estimates that OpenAI’s 2025 revenue will exceed $20 billion, but losses will still reach $14–15 billion. To meet profit expectations for going public, OpenAI must allocate its limited resources to the most commercially valuable businesses.

Right now, enterprise services are the most reliable cash cow in the AI industry. Anthropic’s annual revenue exceeds $19 billion, with about 80% coming from enterprise clients—creating direct competitive pressure for OpenAI.

Shutting down Sora and doubling down on GPT-6 is essentially OpenAI reshaping its commercial narrative before going public.

It’s shifting from attracting consumer users with flashy generation capabilities to serving enterprise clients with stable, efficient productivity tools—from an AI technology explorer to an AI commercialization enabler. And GPT-6’s debut is the core vehicle for this transformation.

02

The AI Race Isn’t a Single-Thread Sprint

While OpenAI was busy preparing for GPT-6’s release, other global AI players didn’t follow its rhythm. Instead, they began a new round of strategic layout (deployments) in advance.

On April 2, Google DeepMind officially launched the Gemma4 series of large models, fully open-sourced under the Apache 2.0 license, with four models covering all scenarios from edge devices to cloud inference.

The E2B model, with 2 billion parameters, occupies as little as 1.5GB of memory and can run offline on smartphones and Raspberry Pis, yet its performance matches the previous 27-billion-parameter model. The 31-billion-parameter dense model ranks third among open-source models on the Arena AI leaderboard, outperforming competitors with 20 times more parameters.

The open-source community responded immediately. On the day of release, over 1,000 derivative models appeared on HuggingFace. Many edge AI development teams directly replaced their previous model solutions. Google’s official data shows Gemma4 surpassed 400 million downloads within 24 hours of release, with over 100,000 derivative models, sparking a new frenzy in the open-source community.

Google’s move directly contrasts with OpenAI’s approach. While OpenAI uses closed-source GPT-6 to sprint toward AGI, Google is using open-source methods to decentralize AI capabilities to global developers and device manufacturers, seizing control over edge AI entry points.

Gemma4’s core breakthrough is freeing large models from cloud computing constraints. Phones, IoT devices, and cars can now perform full-featured AI inference locally. The future AI competition will no longer be just a parameter race for cloud-based large models—it will also be an ecological battle for edge devices.

Domestic tech giants have also been busy. On the same day Google released Gemma4, Alibaba officially launched Tongyi Qianwen Qwen3.6-plus, which performed impressively on the global authoritative programming benchmark CodeArena. It ranked second globally in the React special project (specialized) ranking list , which evaluates autonomous coding capabilities for complex web development—surpassing OpenAI’s GPT-5.0-high and Google’s Gemini 3.1pro, second only to Anthropic’s Claude-opus-4.6.

ByteDance launched Doubao 2.0 this year, supporting private deployment to meet enterprise data security and compliance needs, while announcing that its Doubao series has joined the top tier of global AI assistants. Industry-leaked data shows Doubao’s daily token consumption has reached 120 trillion, growing 1,000-fold from its initial stage and doubling in three months.

Tencent, in late March, fully rolled out the ClawBot plugin for WeChat, natively integrating the OpenClaw open-source AI agent framework—turning the chat boxes of 1.2 billion users directly into AI consoles capable of autonomously breaking down complex tasks.

Microsoft’s moves are even more intriguing. It recently launched three self-developed commercial models: MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2, covering speech transcription, speech generation, and image generation. Developed by Microsoft’s MAI superintelligence team, these models signal that beyond its partnership with OpenAI, Microsoft is accelerating its own AI model ecosystem to reduce reliance on external technologies.

Across the entire AI industry, a clear strategic divergence has emerged ahead of the wave GPT-6 is about to create.

OpenAI is sprinting toward the AGI finish line along a closed-source path, attempting to redefine industry ceilings with its next-generation model. Google, Alibaba, and others are using open-source approaches to build broader developer ecosystems and seize the future of edge AI. Microsoft, Tencent, and other giants are integrating AI capabilities deeply into existing user scenarios through their product ecosystems, achieving rapid technological deployment.

This AI competition, which has lasted over three years, is no longer a single-thread sprint.

03

Deployment Is More Important Than Parameters

From ChatGPT’s release in late 2022 to now, over three years have seen the AI industry engage in a frenzied parameter arms race. Companies competed to release larger, more precise models—but few truly focused on solving the core issue of how AI could be deployed.

Most large models remain stuck at the chatbot stage. Beyond a few tech companies and developers, the vast majority of ordinary users and traditional industries haven’t truly benefited from AI-driven efficiency gains.

With GPT-6’s arrival, based on current limited information, it seems poised to redefine the race’s finish line. The industry’s competitive focus is shifting from “can we build it?” to “can we use it?”

Whether it’s GPT-6’s deep integration of agent capabilities, Gemma4’s extreme optimization for edge deployment, or Qwen3.6-plus’s refinement for programming scenarios, the essence is moving AI technology out of labs and into real work environments—turning it into tools that solve practical problems.

For domestic large model vendors, GPT-6’s release presents both challenges and fresh opportunities. For years, Chinese models have played catch-up, following overseas vendors’ rhythm (pace) in parameter and precision races. But as the AI industry’s focus shifts to deployment and ecosystems, domestic vendors—with their local scenario advantages, user bases, and industrial chain resources—now have a chance to achieve differentiated breakthroughs.

According to enterprise service practitioners, traditional industries like manufacturing, finance, and retail in China are seeing rapidly growing demand for AI. However, what they need isn’t a do-it-all general-purpose large model but vertical solutions that address industry-specific pain points.

For example, manufacturing needs quality inspection and production flow optimization; finance requires risk control and customer service; retail focuses on user operations and supply chain management. These scenarios demand AI models deeply integrated with industry knowledge—exactly where domestic vendors excel.

For ordinary developers and users, there’s no need to overanxiously worry about whether GPT-6 will disrupt the industry or obsessively chase the latest model versions.

What truly matters is learning to make these AI tools helpers in your work. AI shouldn’t exist to replace humans—it should free people from repetitive, tedious tasks to focus on more valuable work.

GPT-6’s launch event will arrive in a few days. It may indeed refresh our understanding of AI capabilities and bring new changes to the industry.

But we must also recognize that even though Sam Altman once said AGI would be one of the most important events in human civilization, before that distant destination (destination) arrives, the entire industry—and each of us—must first complete a more pragmatic journey.

This article is original to Xinmou

— END —