OpenAI's Code Red & Microsoft's Damage Control

In 2022, Google sounded the alarm with a "code red" after OpenAI launched ChatGPT. Three years later, OpenAI finds itself on the defensive.

The artificial intelligence industry is experiencing a dramatic recalibration. What was supposed to be "the year of the agent" has become a sobering reality check, while an unexpected twist has emerged: Google, once written off as hopelessly behind in the AI race, has surged back to reclaim its position as the frontrunner. The combination of agent technology hitting friction and Google's stunning comeback with Gemini 3 has forced the entire industry, including OpenAI, to fundamentally rethink its strategy.

The Agent Economy Hits a Wall

Microsoft recently found itself in damage control after reports surfaced about lowered AI software targets. While the company clarified that growth targets were adjusted but sales quotas remained unchanged, the distinction reveals something profound: even tech giants are struggling to set realistic expectations for AI agent revenue.

The timing is significant. OpenAI CEO Sam Altman reportedly issued an internal "code red" memo announcing a strategic pivot away from agent development to focus on fixing the core ChatGPT experience. For a company that championed AI agents as the future of work, this represents a major course correction.

New data from Ramp tells the real story. While Microsoft and OpenAI recalibrate their agent strategies, Anthropic recorded one of its biggest monthly gains in enterprise AI adoption. But the growth isn't coming from agents or chatbots, it's being driven by API spending. This distinction matters enormously: API usage represents bottom-up adoption by developers solving specific problems, while enterprise agents are sold top-down to executives with promises of sweeping transformation.

Google's Stunning Comeback

Just 18 months ago, critics were calling for Google CEO Sundar Pichai to step down. The company seemed hopelessly behind, unable to catch up and dominate AI. Then came Gemini 3 Pro.

Google's latest model doesn't just compete with OpenAI's GPT, it beats it across nearly every benchmark. But more importantly, Google has revealed itself as the most comprehensively positioned company in AI. The company possesses every critical advantage simultaneously: frontier models, massive AI infrastructure, custom TPU silicon, diversified revenue streams, top researchers, consumer hardware dominance (Android, Pixel), billions of users, proprietary data across Search, YouTube, Gmail, and Maps, and the ability to deeply integrate AI into products people use daily.

Google's stock price surged over 15% in a single month, adding hundreds of billions in market cap to an already $3 trillion company. Investors finally recognized what had been building: Google didn't just catch up, they positioned themselves to win.

OpenAI's Code Red Response

Sam Altman's leaked memo to OpenAI staff acknowledged the threat directly: "Google will create some temporary economic headwinds for our company." But he remained defiant, noting that "ChatGPT is AI to most people and I expect that to continue." The memo revealed OpenAI's existential challenge: "It sucks that we have to do so many hard things at the same time. The best research lab, the best AI infrastructure company, and the best AI platform product company."

This statement captures OpenAI's fundamental vulnerability. Unlike Google, which can fund AI investments from diversified revenue streams, OpenAI must raise financing for every data center buildout. As Meta CEO Mark Zuckerberg recently noted, his company is willing to "misspend potentially hundreds of billions of dollars" on AI because they won't go bankrupt even if wrong. OpenAI lacks that safety net, every major infrastructure bet carries existential risk.

The End of the Scaling Era

Ilya Sutskever, OpenAI co-founder and current head of Safe Superintelligence, offers a compelling framework for this moment. According to Sutskever, we're transitioning from "the age of scaling" back to "the age of research."

From 2020 to 2025, the industry operated under a simple mantra: scale everything. Pre-training was the recipe, more data plus more compute yielded better results. But now, as Sutskever observes, "we are in a world where there are more companies than ideas by quite a bit."

Pre-training is running out of data, and the path forward isn't obvious. According to reports, OpenAI's researchers haven't completed a successful full-scale pre-training run broadly deployed for a frontier model since GPT-4o in May 2024, over 18 months ago. This highlights the significant technical hurdle Google's TPU fleet managed to overcome.

The Generalization Problem

At the heart of AI's limitations lies what Sutskever identifies as the fundamental challenge: models "generalize dramatically worse than people." He illustrates this with coding AI that fixes one bug by introducing another, then brings back the first bug when you point out the second, creating an endless loop.

Sutskever suggests companies may be inadvertently "reward hacking" by designing reinforcement learning environments inspired by the benchmarks they want to excel at. This produces models that ace evaluations but fail in messy real-world scenarios, explaining the disconnect between impressive benchmark scores and frustrating practical performance.

Sutskever uses a human analogy to compare current models to a student who spends 10,000 hours honing competitive programming skills to win contests, versus a naturally gifted student who practices for just 100 hours. While the former may excel in competitions, the latter, with that elusive "it" factor, is likely to have the more successful career.

OpenAI Fights Back with Pre-Training Focus

Despite the challenges, OpenAI isn't conceding defeat. In a striking reversal, the company is pivoting back to pre-training fundamentals. Mark Chen, OpenAI's chief research officer, recently stated they've been "supercharging our pre-training efforts for the last half year." He explicitly mentioned Google by name: "We can go head-to-head with Gemini 3 easily on pre-training."

This represents a significant strategic shift. While the industry spent recent years focused on reinforcement learning and reasoning models, OpenAI acknowledged it had "muscle atrophy on the pre-training side." Now they're going all-in on pre-training to directly counter Google's advances.

According to The Information, OpenAI is developing "Garlic," a new model specifically designed to counter Google's gains through enhanced pre-training. Early internal evaluations show it performing well against both Gemini 3 and Anthropic's Claude Opus 4.5. Additionally, OpenAI has a new reasoning model "ahead of Gemini 3" in internal testing, potentially launching in the next few months.

As part of the code red initiative, OpenAI is pausing projects like advertising and shopping features to focus entirely on improving ChatGPT's core performance, personalization, speed, and reliability.

The Real Battleground

Interestingly, Chen's comments reveal an important insight: "For 99% of use cases, we don't really need higher intelligence models." What matters most is the day-to-day experience, the scaffolding around the models. This explains why ChatGPT remains dominant despite Google's technical achievements. ChatGPT is a verb, like "Google it" before it. People say "I'll ChatGPT that." Brand loyalty and user experience matter more than benchmark scores for most consumers.

Google has grown from 450 million to 650 million active users in just months, while OpenAI approaches one billion. But OpenAI's growth is slowing, and Google has distribution advantages through Search and Android that OpenAI cannot match.

The Strategic Landscape

The competitive dynamics reveal stark differences in positioning. Google stands alone in having every critical element: frontier models, infrastructure, custom silicon, diversified revenue, proprietary data, billions of users, and deep product integration.

Microsoft finds itself in an interesting position. The company has spent years building its enterprise AI strategy around OpenAI through a major investment partnership. However, Microsoft is also reported to be investing in Anthropic, suggesting a hedge against over-reliance on any single AI partner in a multi-model future.

Amazon's AWS, meanwhile, has been Anthropic's primary cloud and training partner for years. With Claude's enterprise adoption accelerating through API usage, the bottom-up developer adoption that's actually generating revenue, Amazon's early bet on Anthropic looks increasingly strategic.

What Comes Next

Sutskever's vision challenges conventional thinking about Artificial General Intelligence (AGI). Rather than a pre-trained system that knows everything, he advocates for continual learning agents, superintelligent learners that rapidly acquire new skills but start with limited knowledge, like a highly capable 15-year-old.

This approach acknowledges that even humans aren't "AGI" in how the term is commonly used. We rely on continual learning throughout our lives. A superintelligent system might function similarly: deployed into organizations where it learns specific roles, much like a human worker joining a company.

His timeline? Five to 20 years, a notable departure from the aggressive predictions dominating industry discourse just a year ago.

Implications for the Industry

Several critical insights emerge from this inflection point:

• The agent timeline has shifted dramatically. What was promised in months will take years, potentially decades. The gap between demo capabilities and production reliability is wider than anticipated.
• Infrastructure and capital matter more than ever. Companies with diversified revenue streams can make multiple big bets without existential risk. Pure AI plays like OpenAI must be right on fewer attempts.
• Distribution is destiny. Google's integration across Search, Android, YouTube, and Gmail provides advantages no pure AI company can match. Apple's potential entry could be similarly disruptive if they finally leverage their hardware ecosystem.
• Pre-training isn't dead, but it's harder. Google proved scaling works with the right architecture (TPUs) and approach. OpenAI is now betting heavily on catching up through renewed pre-training focus. But the low-hanging fruit is gone, breakthroughs require both massive compute and novel ideas.
• The real money flows to APIs, not agents. Developers integrating AI into existing workflows generate revenue today. Autonomous agents remain a promise for tomorrow.

The Bottom Line

AI is transforming work, but gradually through APIs and model integration, not a sudden agent-driven revolution. The hype cycle has met reality.

Google is the clear leader, with every piece of the puzzle. OpenAI remains strong due to brand dominance and user loyalty but faces challenges around infrastructure and capital. Its return to pre-training fundamentals shows adaptation, though it lags behind Google in this area.

The shift from scaling to research increases uncertainty. Progress now relies on unpredictable breakthroughs, meaning AI returns will come, but over a longer and more uncertain timeline than expected.

Consumers benefit from fierce competition between Google, OpenAI, and others, driving innovation despite longer timelines. The AI revolution isn't cancelled; it's maturing into a phase where sustainable business models are as important as model performance.