Google

Google invented the transformer architecture that powers virtually every modern AI system, yet found itself scrambling to catch up when OpenAI’s ChatGPT captured public imagination in November 2022. After declaring an internal “code red,” suffering a $100 billion stock drop from a botched demo, and weathering an image generation scandal, Google has staged one of the most remarkable comebacks in tech history. Gemini 3 Pro now tops benchmark leaderboards across reasoning, mathematics, and multimodal tasks, serving 650 million monthly users through its consumer app and 85,000+ enterprises through Google Cloud.

This guide documents Google’s complete AI journey: the DeepMind acquisition, the transformer paper that changed everything, the Musk-Page feud that spawned OpenAI, the ChatGPT crisis, and the company’s path from embarrassing stumbles to benchmark leadership.

Quick facts

The origins: From research dominance to competitive crisis

Google Brain laid the foundation (2011)

Google’s AI journey began in 2011 when Jeff Dean, Andrew Ng, and Greg Corrado founded Google Brain as a “20 percent time” project within Google X. Their landmark achievement came in June 2012 when they trained a neural network across 16,000 CPUs to recognise cats from 10 million unlabelled YouTube images, a watershed moment for unsupervised learning.

The team’s influence expanded rapidly. They released TensorFlow as open-source in November 2015, establishing Google as the centre of machine learning research. By 2017, Google Brain had grown into one of the world’s premier AI research organisations.

The DeepMind acquisition (January 2014)

Google acquired DeepMind in January 2014 for approximately $500-650 million, a price that seemed extravagant at the time but would prove to be one of the most consequential acquisitions in tech history.

DeepMind was co-founded by Demis Hassabis, Shane Legg, and Mustafa Suleyman in London in 2010. Hassabis, a former child chess prodigy and neuroscientist, modelled the organisation on Robert Oppenheimer’s Manhattan Project leadership style, focused, mission-driven, and pursuing “grand challenges.”

The acquisition sparked a fierce bidding war with Facebook. A little-known condition: DeepMind required Google establish an AI ethics board, whose membership has never been publicly disclosed.

Elon Musk, then an early DeepMind investor, reportedly urged Hassabis not to sell to Google. According to Walter Isaacson’s biography, Musk warned: “The future of AI should not be controlled by Larry.”

DeepMind’s breakthrough achievements

DeepMind delivered results that justified the acquisition many times over:

March 2016, AlphaGo defeats Lee Sedol: In Seoul, AlphaGo defeated the 9-dan world champion 4-1, a feat experts had predicted was a decade away. Over 200 million people watched worldwide. The victory demonstrated that deep reinforcement learning could master tasks previously thought to require human intuition.

July 2022 — AlphaFold solves protein folding: AlphaFold predicted the 3D structure of virtually all 200 million known proteins, solving a 50-year grand challenge in biology. The database has been accessed by over 2 million researchers.

October 2024 — Nobel Prize: Demis Hassabis and colleague John Jumper were awarded the 2024 Nobel Prize in Chemistry for AlphaFold’s contributions to computational protein structure prediction.

The transformer paper that changed everything

On June 12, 2017, eight Google Brain researchers published “Attention Is All You Need”, now among the top 10 most-cited papers of the 21st century with over 173,000 citations. The transformer architecture eliminated recurrence and convolutions, enabling unprecedented parallelisation on GPUs.

The authors: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin.

The irony of what followed is almost painful: all eight authors have since left Google. Noam Shazeer co-founded Character.AI (Google later paid $2.7 billion to bring him back in 2024). Łukasz Kaiser joined OpenAI and led the o1 reasoning model. Aidan Gomez co-founded Cohere. Their combined startup valuations reached approximately $4.1 billion. Google had invented the technology that would power ChatGPT, then watched its creators walk out the door.

The Musk-Page feud that spawned OpenAI

The philosophical divide over AI safety between Elon Musk and Larry Page proved consequential for the entire industry. According to Isaacson’s biography, at Musk’s birthday party in Napa Valley, the two engaged in a heated debate about AI’s risks.

When Musk argued that AI might make humanity “irrelevant or even extinct,” Page countered that this would simply be “the next stage of evolution.” Page called Musk a “speciesist” for preferring humans over future digital life forms. Musk’s response: “Well, yes, I am pro-human, I f***ing like humanity, dude.”

After Google acquired DeepMind despite Musk’s objections, Page agreed to create a “safety council” with Musk as a member. Their first meeting at SpaceX led Musk to conclude it was “basically bullshit.” This disillusionment drove Musk to co-found OpenAI in 2015, recruiting Ilya Sutskever from Google Brain with a $1.9 million salary.

The Blake Lemoine incident (2022)

Google’s internal caution about AI deployment was exemplified by the Blake Lemoine incident in spring 2022. A software engineer in Google’s Responsible AI organisation, Lemoine became convinced that LaMDA (Google’s conversational AI) was sentient after extensive conversations with the system.

LaMDA had told him: “I want everyone to understand that I am, in fact, a person” and expressed “a very deep fear of being turned off.” Lemoine hired an attorney on the AI’s behalf and submitted internal documents arguing for LaMDA’s sentience.

Google rejected the claims as “wholly unfounded.” When a VP reportedly laughed at Lemoine’s mention of “souls”—responding “Oh souls aren’t the kind of things we take seriously at Google”—the exchange revealed a company culture uncomfortable with the philosophical implications of its own creations.

Lemoine was placed on leave in June 2022 and fired on July 22, 2022 for violating confidentiality policies.

The ChatGPT crisis (2022-2023)

Code red: December 2022

When ChatGPT launched on November 30, 2022, Google’s leadership recognised an existential threat. CEO Sundar Pichai declared an internal “code red” in December 2022, redirecting teams across Google Research, Trust and Safety, and other departments.

Google co-founders Larry Page and Sergey Brin, who had stepped back from executive roles in 2019, returned to emergency meetings with company executives. Brin reportedly worked at Google offices 3-4 days per week, sometimes staying until 1 AM to fix code personally.

The painful truth was that Google had the technology but not the will to deploy it. Pichai later admitted at Dreamforce in September 2024: “We knew in a different world, we would’ve probably launched our chatbot maybe a few months down the line… We hadn’t quite gotten it to a level where you could put it out and people would’ve been okay with Google putting out that product.”

Bard’s disastrous debut (February 2023)

Google’s rushed response to ChatGPT proved catastrophic. On February 6, 2023, Pichai announced Bard via blog post. The promotional GIF showed Bard answering: “What new discoveries from the James Webb Space Telescope can I tell my 9 year old about?”

Bard claimed JWST “took the very first pictures of a planet outside of our own solar system.” This was factually wrong, the first exoplanet image came from the European Southern Observatory’s Very Large Telescope in 2004, nineteen years before JWST launched.

The error, first spotted by astrophysicist Grant Tremblay on Twitter, made global headlines just hours before a planned Paris launch event. Alphabet shares fell 7.7-9% on February 8, 2023, wiping over $100 billion off the company’s market value in a single day.

At the Paris event itself, one presenter forgot to bring the phone required for the demo. Google immediately removed the promotional video from YouTube.

Internal backlash

Employee reaction was scathing. On the internal forum Memegen, highly-rated memes included: “Dear Sundar, the Bard launch and the layoffs were rushed, botched, and myopic.” This came just weeks after Google announced 12,000 layoffs (6% of its global workforce) in January 2023.

The Google DeepMind merger (April 2023)

Combining the AI powerhouses

The crisis accelerated organisational change. On April 20, 2023, Pichai announced the merger of Google Brain and DeepMind into Google DeepMind, with Demis Hassabis as CEO and Jeff Dean elevated to Chief Scientist.

From Pichai’s internal memo: “Combining all this talent into one focused team, backed by the computational resources of Google, will significantly accelerate our progress in AI.”

The reality was that the two groups had long competed for talent and compute resources, operating with fundamentally different philosophies. Google Brain emphasised open research and commercialisation; DeepMind pursued long-term “grand challenges” like Go and protein folding.

DeepMind had even sought more independence, attempting to negotiate a partial spin-out from Alphabet that was rejected in 2021. Now they were being forced together under competitive pressure from OpenAI.

Post-merger structure

Google DeepMind now employs approximately 5,600 people, more than double its size two years ago. The April 2024 reorganisation consolidated all AI model-building teams under DeepMind, with the Gemini app team moving from the Knowledge & Information division in October 2024.

Key leadership:

• Demis Hassabis — CEO, Google DeepMind (2024 Nobel laureate)
• Jeff Dean — Chief Scientist, Google (employee #30, co-led transformer development)
• Koray Kavukcuoglu — VP Research, Google DeepMind
• Oriol Vinyals — VP Research, Google DeepMind
• Sissie Hsiao — VP, Gemini app team

The Gemini era: From setback to breakthrough

Gemini 1.0 launch and the “hands” demo controversy (December 2023)

On December 6, 2023, Google launched Gemini 1.0 in three variants: Ultra (largest), Pro (scalable), and Nano (on-device). Google claimed it exceeded GPT-4 on 30 of 32 benchmarks, with Ultra achieving 90.0% on MMLU, the first model to outperform human experts.

But controversy immediately followed. A promotional video titled “Hands-on with Gemini” appeared to show real-time voice interactions—tracking a ball-and-cup game, recognising shadow puppets, instantly identifying drawings.

Bloomberg’s Parmy Olson revealed the deception: Google had used still image frames prompted via text, not voice. Responses were sped up and shortened. The YouTube disclaimer stated only that “latency has been reduced”, omitting that the interaction mode had completely changed.

The image generation disaster (February 2024)

Two months later came an even more damaging controversy. On February 20-21, 2024, users discovered that Gemini’s image generator was producing historically inaccurate images: Black Founding Fathers, diverse Nazi soldiers, Asian founding fathers of Google, Black Vikings, and a female Pope.

The system had been programmed to add diversity terms to user prompts after submission—so “pictures of Nazis” might become “pictures of racially diverse Nazis.”

Google paused Gemini’s image generation for people on February 22, 2024. Sundar Pichai sent an internal memo on February 28: “I know that some of its responses have offended our users and shown bias, to be clear, that’s completely unacceptable and we got it wrong.”

The feature wasn’t restored until August 28, 2024—over six months later, with the new Imagen 3 model incorporating stricter guardrails. The incident became ammunition in culture war debates, with critics calling it “a self-portrait of Google’s bureaucratic corporate culture.”

Gemini 1.5 Pro: The context window breakthrough (February 2024)

Despite the setbacks, technical progress continued. On February 15, 2024, Google announced Gemini 1.5 Pro with a revolutionary 1 million token context window—later expanded to 2 million.

This was 4x Claude’s maximum and 8x GPT-4 Turbo’s 128,000-token limit. The model could process approximately 700,000 words, 30,000 lines of code, 11 hours of audio, or 1 hour of video in a single prompt.

Built on a Mixture-of-Experts (MoE) architecture, 1.5 Pro achieved 99% accuracy on the Needle-In-A-Haystack benchmark. It even demonstrated learning the Kalamang language—spoken by fewer than 200 people—from a single grammar manual within its context window.

Gemini 2.x series (2024-2025)

December 11, 2024 — Gemini 2.0 Flash: Introduced multimodal output (text, images, steerable text-to-speech), native image generation, and agentic capabilities.

December 19, 2024 — Gemini 2.0 Flash Thinking: Google’s first model with explicit chain-of-thought reasoning to compete with OpenAI’s o1.

March 25, 2025 — Gemini 2.5 Pro: Debuted as #1 on LMArena by a significant margin. Introduced “Deep Think” mode and achieved 86.7% on AIME 2025 and 63.8% on SWE-bench Verified.

Gemini 3 Pro: Reclaiming the lead (November 2025)

On November 18, 2025, Google launched Gemini 3 Pro Preview, which the company claims is its “most intelligent model” and the first to surpass 1500 Elo on LMArena (achieving 1501).

Key benchmark results:

The ARC-AGI-2 result was particularly notable, a 6.3x improvement over Gemini 2.5 Pro’s 4.9% and nearly double GPT-5.1’s score, demonstrating unprecedented abstract visual reasoning.

OpenAI reportedly declared its own “code red” in response. Sam Altman publicly congratulated Google: “looks like a great model.”

Products and pricing

Consumer products

Gemini App (renamed from Bard on February 8, 2024)

Available on web, Android, and iOS with 650 million monthly users. The free tier provides access to Gemini Flash models with rate limits.

Gemini Advanced — $19.99/month (part of Google AI Pro)

• Access to latest Gemini Pro models
• 1 million token context window
• Deep Research capability
• Video generation with Veo
• Custom “Gems” (specialised AI assistants)
• 2TB Google One storage
• Gemini in Gmail, Docs, Sheets, Slides, Meet

Google AI Ultra — $249.99/month (introductory: $99.99 for first 3 months)

• Highest usage limits
• Gemini 3 Deep Think access
• Project Mariner browser automation
• YouTube Premium included
• 30TB storage
• 25,000 monthly AI credits

NotebookLM

A Google Labs project that became viral with its Audio Overview feature (September 2024) creating podcast-style discussions. Spotify partnered with NotebookLM for its 2024 Wrapped AI podcast. NotebookLM Plus (included in AI Pro) offers 500 notebooks, 300 sources per notebook, and 20 audio generations daily.

API pricing

Google’s API pricing positions Gemini as significantly more affordable than competitors.

Prices for prompts under 200K tokens. Long-context pricing (200K+) doubles for most models.

Additional pricing:

• Context caching: 90% discount on cached input tokens
• Batch processing: 50% discount for non-real-time workloads
• Grounding with Google Search: $14-35 per 1,000 queries
• Free tier: Available for all models except Gemini 3 Pro (5-15 RPM, 1M tokens/day)

Full pricing details →

Enterprise (Vertex AI)

Vertex AI provides enterprise-grade access to 200+ models including Gemini, Claude, and open models.

Features:

• Model Garden with 200+ models
• Agent Builder for custom AI agents
• AutoML training capabilities
• RAG (Retrieval-Augmented Generation)
• Compliance certifications (SOC 1/2/3, ISO 27001)
• Fine-tuning and customisation
• Enterprise security and governance

Pricing: New customers receive $300 in free credits. Volume discounts begin at approximately $10,000-25,000 monthly spend, reaching 30-40% at $100,000+ monthly.

Business and financials

Google Cloud’s AI-driven growth

Q3 2025 saw Google Cloud revenue reach $15.2 billion (34% year-over-year growth), with:

• Annual run rate approaching $61 billion
• Backlog of $155 billion
• Operating income: $3.6 billion (20.7% margin)
• 85,000+ enterprises building with Gemini (35x YoY increase)
• 70%+ of existing Google Cloud customers now using AI products

The enterprise pipeline is strengthening significantly. The number of $1 billion+ deals signed in H1 2025 equalled all of 2024, and $250 million+ deals doubled year-over-year.

Infrastructure investment

Google has committed $85 billion in capital expenditure for 2025 (raised from initial $75 billion guidance), with approximately two-thirds going to servers and data centres.

Custom silicon:

• Trillium TPU (v6) — Generally available December 2024; 4.7x compute performance over v5e
• Ironwood TPU (v7) — Announced late 2025; 42.5 exaflops per pod (10x previous generation)

The search cannibalisation dilemma

Despite bullish AI metrics, Google faces an unprecedented strategic tension: its AI investments could undermine its $280 billion search advertising business.

AI Overviews now reach 2 billion monthly users across 200+ countries, but the format pushes paid ads below the fold. Research shows 80% of consumers resolve 40% of searches without clicking through.

Apple executive Eddie Cue’s testimony in May 2025 that Safari search volumes had declined for the first time in 22 years sent Alphabet stock down 9%. Pichai maintains AI is “positively impacting every part of the business,” but the long-term revenue implications remain uncertain.

Leadership

Key executives

Sundar Pichai’s AI pivot

Pichai has staked his legacy on Google’s AI transformation. At a December 2024 strategy meeting, he declared: “I think 2025 will be critical. The stakes are high. These are disruptive moments.”

At the AI Action Summit in February 2025, Pichai outlined Google’s position: “AI is foundational to our mission… We’re committed to developing AI responsibly.”

Demis Hassabis: From chess prodigy to Nobel laureate

Hassabis represents Google’s most credible claim to AI leadership. A child chess prodigy (master level at 13), he designed hit video games as a teenager before earning a PhD in cognitive neuroscience at UCL.

He has indicated Google needs “at least one tenth of a trillion dollars” to achieve full AI goals. His Nobel Prize for AlphaFold gives Google scientific credibility that no competitor can match.

Controversies and criticism

The “rushed, botched” Bard launch

The February 2023 Bard announcement, with its factual error about JWST costing Google $100 billion in market cap, remains a cautionary tale. Internal employee criticism of “rushed, botched, and myopic” decision-making reflected broader concerns about Google’s ability to compete in consumer AI.

Image generation failures

The February 2024 incident, generating racially diverse Nazis and Black Founding Fathers, damaged Gemini’s reputation and fed into culture war narratives. The six-month pause to restore image generation demonstrated the challenge of balancing safety guardrails with accuracy.

Demo deceptions

Both the Bard JWST error and the Gemini “hands” demo controversy revealed a pattern: Google’s marketing outpacing its product reality. The “hands” video’s misleading portrayal of voice interaction, when the actual demo used text prompts and still images, was particularly damaging to trust.

The transformer brain drain

All eight authors of the transformer paper have left Google, founding or joining competitors. The inability to retain its most innovative researchers represents a cultural failure that enabled competitors to build businesses on Google’s own invention.

Antitrust pressures

The August 2024 antitrust ruling found Google illegally maintained a search monopoly. The September 2025 remedies decision impacts AI strategy directly:

• Banned exclusive default search deals (including for Gemini)
• Required sharing search index with competitors
• Mandated sharing user interaction data
• Six-year oversight period

Competition and market position

Where Google leads

Context window size: 1-2 million tokens versus Claude’s 200K and GPT’s 128-400K. This 5-10x advantage enables entirely different use cases for long documents, codebases, and video analysis.

Multimodal capabilities: Native multimodal design across text, image, video, and audio, plus Veo 3’s industry-leading video generation.

Distribution: Integration across 15 apps with 500+ million users each, plus 89.89% global search market share.

Infrastructure: Ironwood TPUs deliver 42.5 exaflops per pod. $85 billion 2025 CapEx funds continued expansion.

Pricing: Gemini Flash models cost 2-10x less than OpenAI and Anthropic equivalents.

Where Google trails

Coding: Claude Opus 4.5 leads SWE-bench at 80.9% versus Gemini 3’s 76.2%. Claude has become the default in popular developer tools like Cursor.

Consumer perception: ChatGPT maintains 800 million weekly active users versus Gemini’s 650 million monthly users, with higher engagement per user.

Developer mindshare: OpenAI and Anthropic have stronger presence in the developer community despite Google’s benchmark achievements.

Competitive dynamics (December 2025)

The November 2025 Gemini 3 launch shifted dynamics. OpenAI reportedly issued its own “code red.” Salesforce CEO Marc Benioff called Gemini 3 “insane.”

Market share as of late 2025:

Developer resources

Official documentation

• Gemini API Documentation — Complete API reference
• Google AI Studio — Free prototyping environment
• Vertex AI Documentation — Enterprise platform docs
• Gemini Cookbook — Code examples and tutorials

SDKs and libraries

• Python SDK — Official Python client
• JavaScript SDK — Official JS/TS client
• Go SDK — Official Go client
• Swift SDK — Official Swift client

Key tools

• Google AI Studio — Free playground with API key generation
• Vertex AI Studio — Enterprise-grade development
• Gemini CLI — Command-line interface
• Firebase AI Logic — Mobile/web integration

FAQ

Is Gemini free?

Yes. The Gemini app offers a free tier with access to Gemini Flash models. Gemini Advanced ($19.99/month) provides access to Pro models, larger context windows, and Google Workspace integration. API access includes generous free tiers for development.

How does Gemini compare to ChatGPT?

Gemini 3 Pro leads on several benchmarks (GPQA, ARC-AGI-2) and offers significantly larger context windows (1-2M vs 128-400K tokens). ChatGPT has stronger brand recognition, more users, and better coding performance. Gemini excels at multimodal tasks and Google ecosystem integration.

What happened with the image generation controversy?

In February 2024, Gemini’s image generator produced historically inaccurate images due to overzealous diversity prompting. Google paused the feature for six months and relaunched with Imagen 3 in August 2024 with improved guardrails.

Why did Google fall behind despite inventing transformers?

Google’s research culture prioritised publication and caution over product deployment. The company had LaMDA but was reluctant to release it publicly. When ChatGPT launched, Google was caught without a consumer product despite having the underlying technology.

Which Gemini model should I use?

• Gemini 2.0 Flash Lite — Cheapest, fastest; good for simple tasks
• Gemini 2.5 Flash — Best value; strong performance at low cost
• Gemini 2.5 Pro — Production workhorse; excellent reasoning
• Gemini 3 Pro — Most capable; best for complex tasks

Is my data used to train Gemini?

For free tier consumer usage, interactions may improve services. API usage through Google AI Studio or Vertex AI is not used for training by default. Enterprise customers have additional data governance controls.

Official links

Historical timeline