For most of 2025, it looked like Google had finally figured out the AI game. Gemini was climbing benchmarks, gaining market share, and for the first time in years, people were genuinely excited about a Google AI product. Then something shifted. Not one dramatic failure, but a series of self-inflicted wounds that together paint a picture of a company struggling to convert technical excellence into sustained leadership. Google is still innovating, sometimes brilliantly. But innovation alone isn't winning, and the gap between Google and its competitors is widening in the places that matter most.

The Antigravity debacle

Google Antigravity launched as an ambitious agent-first IDE, designed to evolve the code editor into something fundamentally new. Early reviews were glowing. Developers praised its planning mode, its browser agent that could navigate and test UIs autonomously, and the deep integration with Gemini 3 Pro. For a brief window, it genuinely felt like the best AI coding tool available. Then Google pulled the rug. Paid Pro users started hitting rate limits after just 20 minutes of usage. Worse, once you hit the cap, you weren't waiting a few hours for a reset. You were locked out for seven days. Google had silently rolled out new "agentic" features while slashing quotas, with no email, no notification, and no acknowledgment on support forums. Users on Reddit called it a "textbook bait-and-switch." Some filed complaints under EU consumer protection directives. Others simply cancelled. Google eventually responded by restructuring limits, giving Pro and Ultra subscribers priority access with five-hour refresh cycles, and moving free users to weekly quotas. But the damage was done. Developers had already started migrating to Cursor, Claude Code, and other tools that offered more predictable, transparent usage terms. The deeper problem isn't the rate limits themselves. Every AI tool has constraints. The problem is the pattern: launch with generous terms to capture mindshare, then quietly restrict once users are invested. It erodes trust, and trust is the one thing you can't recover with a better model.

Fragmentation from within

A Bloomberg report from April 2026 revealed what many suspected: Google's AI coding capabilities are scattered across more than six products under different brands, managed by different teams with different priorities. Gemini in the consumer app, Gemini in Google AI Studio, Gemini in Antigravity, Gemini in Vertex AI, coding features in Google Cloud, and various internal tools, all competing for resources and attention. The result is that no single product gets the full weight of Google's engineering talent. Meanwhile, Anthropic has Claude Code, a single, focused product that does one thing extremely well. It's terminal-native, deeply agentic, and designed for developers who want to hand off complex tasks and review the output. Rakuten's engineering team used Claude Code to implement a specific activation vector extraction method in vLLM, a codebase with 12.5 million lines of code, and it finished the entire job in seven hours of autonomous work. Perhaps the most telling detail: Business Insider reported that some Google engineers themselves prefer Claude Code over Google's own tools. An internal divide has emerged between teams that have access to Claude and those that don't, creating what the report described as "the Claude haves and have-nots." When your own engineers are reaching for the competitor's product, that's not a benchmarking problem. That's a product problem.

The market share illusion

On paper, Google's numbers look encouraging. Gemini's share of global GenAI chatbot traffic climbed from 5.7% to over 21% between early 2025 and early 2026, while ChatGPT dropped from nearly 87% to around 65%. Gemini attracted roughly 368 million unique visitors in March 2026. These are real numbers. But context matters. Much of Gemini's growth is distribution-driven, not product-driven. Google bundles Gemini with Google One subscriptions that include 2TB of cloud storage. Users on Reddit openly admit they chose Gemini because the storage bundle made it a no-brainer, not because the AI was better. When your growth strategy is essentially "include it free with something people already pay for," the engagement depth is questionable. Meanwhile, Anthropic went from $1 billion to $14 billion in annualized revenue in roughly 13 months, one of the fastest revenue ramps in software history, driven primarily by enterprise API usage and Claude Code's $2.5 billion annualized run-rate. That's not bundled distribution. That's developers and companies choosing to pay specifically for the AI product.

Where Google is genuinely ahead

It would be unfair to pretend Google isn't doing impressive work. In some areas, they're genuinely leading. PolarQuant and TurboQuant represent a real breakthrough in model compression. PolarQuant transforms key-value cache embeddings into polar coordinates using an efficient recursive algorithm, enabling extreme compression of the memory footprint during inference. Combined with Quantized Johnson-Lindenstrauss projections, TurboQuant achieves roughly 6x memory reduction with effectively zero accuracy loss. This isn't incremental. If it scales as promised, it could fundamentally change the economics of running large language models. Gemini Embedding 2 is Google's first natively multimodal embedding model, mapping text, images, video, audio, and documents into a single embedding space. Unlike the CLIP family of models where cross-modal interaction only happens at the output layer, Gemini Embedding 2 processes all modalities through the same architecture. This matters for retrieval-augmented generation, semantic search, and any application that needs to reason across different types of media. On reasoning benchmarks, Gemini 3.1 Pro scored 94.3% on GPQA, the highest of any mainstream model. Gemini 3 broke 1500 on the LMArena Elo rating. The model quality is legitimately strong, and the one-million-token context window remains an advantage that competitors haven't matched.

The real problem

Google's challenge isn't technical capability. It's organizational execution. They have the best infrastructure, the deepest talent pool, and arguably the strongest research output of any AI lab. They invented the transformer architecture. They have distribution through Android, Chrome, Search, Gmail, and Google Cloud. They have more compute than almost anyone. And yet they keep losing the product race. The pattern repeats: launch something promising, fragment it across teams, confuse the market with multiple overlapping products, alienate early adopters with policy changes, and watch as a more focused competitor captures the high-value users. It happened with Bard versus ChatGPT. It's happening now with Antigravity versus Claude Code. Anthropic ships one coding product and iterates relentlessly. OpenAI maintains a clear product hierarchy with ChatGPT at the center. Google launches six things, argues internally about which one matters, and ends up with none of them being the obvious choice.

What would need to change

Google needs to pick a lane and commit. One AI coding product with the full weight of the organization behind it. One consumer AI experience that isn't a bundle add-on. One set of usage policies that don't change without warning. The research output is still world-class. PolarQuant could reshape inference economics. Gemini Embedding 2 could become the foundation for a new generation of multimodal applications. The models themselves are competitive at the frontier. But none of that matters if developers can't trust the platform they're building on. If your IDE locks you out for a week after 20 minutes of paid usage, the benchmark scores are irrelevant. If your coding tools are scattered across six products and your own engineers prefer the competitor, the context window size doesn't matter. Google has all the pieces. They just can't seem to put them together. And every month they spend reorganizing, their competitors are shipping.

References