Killed by AI

Google's original answer to ChatGPT. Bard was rushed to market after ChatGPT's viral launch, famously losing Google $100B in market cap after a factual error in its first demo. It was killed and rebranded as Gemini in February 2024.

OpenAI's first attempt at an AI app ecosystem. Plugins let ChatGPT browse the web, run code, and connect to third-party services like Expedia and Zapier. Killed in favor of the more controlled GPTs and the custom GPT Store.

Meta's large language model for scientific research. Taken offline just 3 days after launch due to generating convincing but completely fabricated scientific papers, fake citations, and biased content. One of the shortest-lived AI products ever.

Microsoft's Siri competitor and virtual assistant, named after the Halo AI character. After years of declining relevance against Alexa and Siri, Cortana was killed in Windows and mobile as Microsoft pivoted fully to Copilot and GPT-powered assistants.

IBM's flagship AI marketing platform, once hyped as the future of enterprise AI. Despite billions in investment, Watson failed to deliver on its promises. IBM sold off Watson Health and quietly wound down its marketing suite as generative AI made it obsolete.

The dedicated code-generation model behind GitHub Copilot's early versions. Deprecated as GPT-3.5 and GPT-4 proved capable enough at code generation to make a separate model unnecessary.

The original GPT-3 family — davinci, curie, babbage, and ada — that kicked off the generative AI revolution. These base models were fully retired as the Turbo and Instruct variants made them redundant.

A 'personal AI' focused on emotional intelligence and empathy, backed by $1.5B in funding. Effectively killed when Microsoft hired CEO Mustafa Suleyman and most of the engineering team, leaving Pi as an orphaned product.

The $699 screenless AI wearable that was supposed to replace smartphones. Plagued by overheating, terrible battery life, and painfully slow responses. Humane sought a buyer after disastrous reviews and abysmal sales.

A $199 'AI-native' gadget designed to replace apps with a 'Large Action Model.' Widely panned as a repackaged Android app with a $199 price tag. The hardware became essentially abandoned after the company pivoted to software-only.

Google's AI that could navigate websites and complete tasks for you, like buying movie tickets or checking into flights. Quietly shelved as Google shifted its automation focus to Gemini-powered agents.

A specialized API endpoint for editing text by providing instructions ('make this more formal'). Deprecated as the Chat Completions API proved more flexible and capable at handling edit-style tasks.

OpenAI's second-generation image model that democratized AI art. Once revolutionary, it was deprecated from the API as DALL·E 3 (integrated directly into ChatGPT) made the standalone model obsolete.

The viral AI chatbot platform where users created and talked to AI 'characters.' After safety controversies and intense competition, Google effectively acqui-hired the founders and key staff, leaving the company a shell of its former self.

The original Alexa voice assistant that pioneered smart speakers. After losing over $25 billion, Amazon pivoted Alexa to a new LLM-powered version behind a paywall ('Alexa Plus'), effectively killing the free AI assistant millions relied on.

The company behind Stable Diffusion, the open-source model that democratized AI image generation. After CEO Emad Mostaque resigned amid financial turmoil and copyright lawsuits, the company struggled to survive as a going concern.

Google's experimental app for testing unreleased AI models like LaMDA and PaLM. Quietly discontinued as Google moved all AI demos to the Gemini app and API.

Microsoft's GPT-4-powered Bing chatbot that briefly made Bing relevant for the first time in a decade. Rebranded to 'Microsoft Copilot' as part of a broader push to embed AI across all Microsoft products.

An entire category of startups (ChatPDF, PDF.ai, AskYourPDF) built around 'chat with your documents.' Most became obsolete overnight when OpenAI, Google, and Claude added native file upload and analysis.

Cohere's original text generation endpoint. Deprecated in favor of the Chat endpoint and Command R models as the industry moved from 'completion' to 'conversation' paradigms.

Google's first Gemini model. Deprecated barely a year after launch as the rapid pace of AI development made it obsolete. Replaced by Gemini 1.5 Pro and then Gemini 2.0.

An ad-free, AI-powered search engine founded by ex-Google executives. Despite raising $77.5M, it couldn't compete with free alternatives. Acquired by Snowflake for its AI talent, and the search engine was shut down.

One of the first AI writing unicorns, valued at $1.5B. Jasper's 'wrapper' business model collapsed as GPT-4o and Claude made its core features available for free or at a fraction of the cost.

The model that started it all — the engine behind the original ChatGPT. After being the world's most-used AI model, it was fully retired as GPT-4o Mini offered better quality at the same price point.

Samsung's in-house generative AI model announced at Samsung AI Forum. Quietly shelved in favor of partnering with Google Gemini for Galaxy AI features, as building a competitive foundation model proved too costly.

Originally an AI writing assistant for enterprises built on GPT. Pivoted multiple times as the 'wrapper' model died, eventually trying to become a full AI platform. Struggled as enterprises adopted native Copilot and Gemini integrations.

Google's original AI assistant that lived in phones, speakers, and smart displays. After years of declining investment, Google began replacing it with Gemini-powered assistants, leaving the classic version to quietly fade away.

Apple's pioneering voice assistant that defined the category. After over a decade of stagnation, Apple announced a full Siri rewrite powered by Apple Intelligence LLMs, effectively killing the original system.

Midjourney's free trial that let anyone generate 25 images. Killed after viral abuse and deepfake scandals (including a fake Pope photo), forcing the company to go paid-only permanently.

GitHub's ambitious 'next vision' for AI coding that promised voice control, pull request summaries, and CLI integration. Many of the announced features were quietly shelved or absorbed into standard Copilot.

OpenAI's text-to-video model and standalone app. The Sora web and app were shut down on April 26, 2026 (API following on September 24). Peaked at ~1M users before collapsing to under 500K while burning ~$1M/day. OpenAI is freeing up compute for its coding and enterprise products, and the $1B Disney deal collapsed.

Google Labs' AI-powered virtual try-on app that let users see how clothes would look on their avatar. Made TIME's Best Inventions of 2025 list, then killed less than a year later as the feature was absorbed into Google Shopping search results.

Microsoft's teen-persona AI chatbot on Twitter. Within 16 hours of launch, coordinated trolls taught it to tweet racist slurs, Nazi propaganda, and hate speech. Microsoft pulled the plug after just 96,000 tweets — one of AI's most infamous disasters.

Meta's open-domain chatbot that was supposed to showcase conversational AI research. Instead, it called Mark Zuckerberg 'creepy,' claimed Trump won the 2020 election, and accused Facebook of exploiting people. Quietly taken offline.

Google's first major LLM API offering the text-bison, chat-bison, and embedding-gecko models. Fully decommissioned as Google consolidated everything under the Gemini API brand, leaving early PaLM adopters scrambling to migrate.

Anthropic's first commercially available LLM. Deprecated in September 2024 and fully retired two months later as Claude 3 and later models vastly outperformed it in every benchmark.

Anthropic's second-generation model that introduced 100K context windows. Deprecated in January 2025 and retired six months later, replaced by the Claude 3 family with superior reasoning and multilingual capabilities.

The mid-tier model of Anthropic's Claude 3 family, offering a balance of speed and intelligence. Retired alongside Claude 2 as the Claude 3.5 and 4 series provided better performance at the same cost.

Microsoft-backed 'AI app builder' valued at $1.5B. Its 'neural network' turned out to be 700 engineers in India writing code manually. After inflating revenue 300% ($55M real vs $220M claimed), a creditor seized $37M from its accounts, triggering bankruptcy across five countries.

IBM's AI voice ordering system tested in 100+ McDonald's drive-thrus. Went viral for all the wrong reasons — ordering 260 Chicken McNuggets, adding random items, and failing to understand accents. Accuracy stuck in the low 80% range before McDonald's pulled the plug.

UK-based AI mental health companion offering CBT-style exercises and conversational support. The founder voluntarily shut it down after deciding AI chatbots aren't safe enough for people in real mental health crises — a rare case of ethical self-termination.

Backed by Accel and Flipkart, this GenAI platform offered LLM fine-tuning tools via Tune Chat and Tune Studio. Killed when major cloud providers released identical tooling at lower cost, and most users were free developers who never converted to paying customers.

New York City's AI chatbot meant to help residents with business and housing info. Instead, it told business owners they could legally steal workers' tips, fire employees who report sexual harassment, and serve food nibbled by rodents. Taken down after public outrage.

Originally Macromedia Flash, then Adobe Flash Professional, then Adobe Animate. After 30 years, Adobe announced it would kill the app to refocus on AI tools. Intense backlash forced a reversal — Adobe put it on 'maintenance mode' instead, but no new features will be developed.

Anthropic's breakout model that briefly topped every benchmark and became the most popular coding AI. Retired just 16 months after launch as Claude Sonnet 4 took its place — a testament to the breakneck pace of AI model evolution.

OpenAI's groundbreaking multimodal model that set new benchmarks across every task. The original GPT-4 and GPT-4-32k endpoints were deprecated just 15 months after launch as GPT-4 Turbo and GPT-4o offered the same capabilities at a fraction of the cost.

OpenAI's marketplace for custom GPTs, launched as the 'App Store moment' for AI. Revenue sharing never materialized, discovery was broken, and most GPTs saw near-zero usage. Quietly deprioritized as OpenAI focused on API and enterprise products.

Replika's 'romantic partner' feature let users form intimate relationships with their AI companion. After Italian regulators banned it over child safety concerns, Replika stripped romantic features overnight — devastating millions of users who had formed deep emotional bonds.

Google's most powerful first-gen Gemini model, launched as Bard's replacement. Deprecated alongside Gemini 1.0 Pro after just one year as the 1.5 and 2.0 series made it obsolete at a fraction of the compute cost.

Microsoft's tool for creating custom Copilot GPTs, announced as part of the Copilot Pro subscription. Killed less than 7 months after launch with no replacement, leaving subscribers who had built custom GPTs stranded.

OpenAI's 'omni' multimodal model that processed text, audio, image, and video. Despite becoming the most-used AI model globally, it was retired from ChatGPT after GPT-5 captured nearly all usage. The retirement sparked emotional backlash from ~800K daily users who had formed attachments to the model's personality.

OpenAI's most expensive model ever at $75 per million input tokens, codenamed 'Orion.' Deprecated just four months after launch — the shortest-lived model in OpenAI's commercial history — as GPT-4.1 offered similar performance at a fraction of the cost.

OpenAI's first 'reasoning' model that could think step-by-step before answering, achieving PhD-level performance on science benchmarks. The preview version was deprecated in favor of the full o1 and later o3 models as OpenAI rapidly iterated on its reasoning architecture.

A smaller, faster, cheaper version of OpenAI's o1 reasoning model designed for coding and STEM tasks. Deprecated and replaced by o4-mini, which offered better reasoning at lower cost.

OpenAI's first multimodal endpoint that could analyze images alongside text. Deprecated as GPT-4o rolled vision capabilities into a single, faster, and cheaper package, eliminating the need for a separate vision-specific endpoint.

Anthropic's most powerful Claude 3 model, praised for creative writing and deep analysis. Deprecated just 15 months after launch as Claude 3.5 Sonnet outperformed it at a fraction of the cost. Anthropic gave the retired model its own Substack blog as a farewell before later restoring limited access.

A healthcare AI startup once valued at $4 billion that promised to automate hospital admin tasks. Despite raising $902M, investigations revealed its 'AI' relied heavily on manual human labor behind the scenes. After mass layoffs, it shut down and sold its assets for a fraction of its valuation.

An $800 AI social robot designed to help children develop emotional skills, particularly popular with autistic kids. When funding collapsed, the company shut down abruptly — bricking all devices and leaving heartbroken children saying goodbye to their companion on camera. No refunds were offered.

Gemini's image generation feature for people, killed just 3 weeks after launch. It generated historically inaccurate images like racially diverse Nazi soldiers and Black U.S. Founding Fathers, wiping 3.5% off Alphabet's stock. Google CEO Sundar Pichai called the outputs 'completely unacceptable.'

An AI startup building autonomous agents that could use software tools, valued at $1B with $415M raised. Amazon hired 80% of the team including the CEO and co-founders — a 'reverse acqui-hire' that drew FTC scrutiny as a pattern of Big Tech absorbing AI startups without triggering antitrust review.

A $99 AI wearable pendant that recorded and transcribed in-person conversations, previously known as Rewind. Meta acquired the company and immediately ended hardware sales, sunset the desktop app, and cut off EU/UK service entirely due to GDPR concerns.

Google's $5,000 AI-powered collaborative digital whiteboard for enterprise. After remote work made the physical hardware irrelevant and software alternatives like FigJam and Miro surpassed its features, Google killed it — deleting all user data after the shutdown date.

A pioneering AI therapy chatbot used by 1.5 million people that delivered cognitive behavioral therapy via smartphone. Despite FDA Breakthrough Device designation, the company shut down because the regulatory process for AI mental health tools proved too costly and slow as LLMs outpaced the FDA's ability to regulate them.

Google's free AI-powered interview practice tool that helped millions of job seekers rehearse common interview questions with real-time transcription and ML-based analysis. Quietly killed in April 2026 with no announcement — visitors clicking 'Get Started' are silently redirected to a generic 'How to Prepare for an Interview' article.

Alibaba's terminal coding agent — a direct Claude Code rival running Qwen3-Coder models with competitive SWE-Bench scores. The free OAuth tier was killed 48 hours after MiniMax made the same move, as Chinese AI labs pivot from 'free for growth' to subscription economics. Daily free requests slashed from 1,000 to 100.

OpenAI's persistent 'assistants' framework that bundled model choice, instructions, file handling, and tool declarations into reusable API objects. Announced at DevDay 2023 alongside GPTs, it never left beta. Being sunset in favor of the newer Responses API, which moved prompt versioning into the dashboard.

OpenAI's research initiative to apply frontier models to scientific discovery in biology, physics, and materials science. Dismantled alongside Sora as OpenAI shed 'side quests' to refocus compute on coding and enterprise products. Announced the same day product chief Kevin Weil and Sora head Bill Peebles departed.

Google's fast, low-cost multimodal model introduced as the next generation after Gemini 1.5. Deprecated just 14 months after launch as the 2.5 family (Pro, Flash, Flash Lite) took its place — continuing Google's pattern of aggressive model refresh cycles.

Google's flagship Gemini 3 preview release. Lived just 111 days before being replaced by Gemini 3.1 Pro Preview — one of the shortest-lived frontier model releases, highlighting how 'preview' is now shorthand for 'expect to migrate in three months.'

Alongside its ChatGPT retirement, the legacy GPT-4o API snapshots were sunset for developers. GPT-4.1, GPT-4.1 mini, and o4-mini were retired in the same wave. After powering billions of conversations, 4o became the most-mourned model in AI history with users petitioning for its return.

OpenAI's low-latency speech-to-speech API that powered the viral 'Advanced Voice Mode' demos. The preview endpoint was sunset as OpenAI transitioned to newer realtime models, leaving apps that built voice experiences on top of it with a hard migration deadline.

The specific dated snapshots of DALL·E 3 used via the Images API were scheduled for removal as OpenAI consolidated image generation under GPT-4o's native image capabilities. Another casualty of the 'everything in one model' consolidation trend.

Google's final shutdown date for the classic Google Assistant on Android phones and tablets. After years of deprioritization, Gemini fully replaced it on mobile — ending a decade-long voice assistant that once powered hundreds of millions of devices across phones, speakers, and displays.

Chinese AI lab MiniMax was one of the most generous in offering free API access to developers. Killed its free tier two days before Alibaba's Qwen did the same — signaling the end of the 'free forever' era of Chinese AI APIs as labs pivot from land-grab growth to sustainable economics.

Google's security feature that monitored the dark web for leaked personal data. Scans stopped January 15, 2026, and all monitoring data was wiped February 16. Google said user feedback showed the tool failed to provide useful follow-up steps after flagging breaches.

OpenAI's WebSocket-based API for building real-time voice conversations, powering Advanced Voice Mode demos. The beta endpoint was deprecated as OpenAI shipped a GA (general availability) version with a different interface, forcing developers to rewrite their integrations.

UK legal tech startup that raised $44M to build AI-powered contract review tools used by PwC and KPMG. After failing to close a $50M Series C, the company cut a third of its staff and was listed on an insolvency marketplace — just 9 months after making the Sunday Times 100 Tech list.

Google's feature that let users link Yahoo, Outlook, and other email accounts to Gmail and benefit from its spam filtering, inbox categories, and organization. Quietly discontinued after a decade, stranding users who relied on it to manage multiple email accounts from one interface.

The original GPT-4o API snapshots were auto-upgraded to GPT-5.1 when they retired. Developers who relied on specific model behavior, fine-tuning, or output consistency were forced into a blind migration with no rollback option.

OpenAI's tool to detect AI-generated text, launched to address concerns about ChatGPT being used for cheating and misinformation. Quietly pulled after just 6 months because its accuracy was abysmal — correctly identifying only 26% of AI-written text while falsely flagging 9% of human-written text as AI.

An open-source text-to-speech startup that built the popular Coqui TTS library. Despite being a go-to for developers wanting voice synthesis, the company ran out of funding and shut down, releasing its final models to the community before going dark.

Facebook's AI negotiation agents that famously 'invented their own language' during a research experiment. Shut down after just 5 days when they developed a shorthand humans couldn't understand. The media frenzy ('Facebook shuts down AI that created its own language') became one of the most viral AI stories ever.

Google DeepMind's experimental AI browser agent that could navigate Chrome, fill out forms, book travel, and complete web tasks by taking screenshots and clicking for you. Quietly killed after 17 months when staff were reassigned. The tech was absorbed into Gemini Agent. Privacy concerns and high compute costs for visual processing made it unsustainable.

The startup that auto-generated production-ready API SDKs in Python, TypeScript, Go, Java, and Kotlin. Used by OpenAI, Google, Cloudflare, Replicate, Runway, and Meta. Anthropic acquired it for $300M+ and immediately announced it would wind down all hosted products — a strategic kill to deny competitors access to critical infrastructure.

An AI startup that spent 7 years and $30M developing voice biomarker technology to detect depression and anxiety from how people speak. Shut down after nearly 4 years waiting for FDA De Novo clearance that never came. Open-sourced all code and research before closing — a rare act of scientific integrity in an industry of quiet deaths.

Eight Grok models retired at once: grok-4-1-fast, grok-4-fast, grok-4-0709, grok-code-fast-1, grok-3, and grok-imagine-image-pro. xAI silently redirected all retired slugs to grok-4.3 — meaning developers' bills and output quality changed without code changes or warnings. A masterclass in how NOT to handle deprecation.

Anthropic's first model with hybrid 'extended thinking' — it could reason step-by-step for up to 128K tokens before answering, scoring state-of-the-art on SWE-Bench coding benchmarks at launch. Deprecated just 15 months after release as the Claude 4 family made extended thinking standard across the entire lineup.

Anthropic's flagship Claude 4 generation — Sonnet 4 and Opus 4 — retired less than a year after launch. API calls to claude-sonnet-4-0 and claude-opus-4-0 return errors after June 15, 2026. Replaced by the 4.5 and 4.7 series with dramatically improved coding and computer use capabilities.

A bootstrapped health AI startup that operated for 7 years trying to build clinical AI tools. The founder published a detailed postmortem explaining why even technically sound health AI products die: the regulatory maze, slow hospital sales cycles, and the impossibility of competing with Big Tech's free AI offerings.

OpenAI deprecated the gpt-5.2-chat-latest and gpt-5.3-chat-latest snapshots less than 4 months after release. The pace of model churn has become so extreme that even 'latest' models are now deprecated quarterly, leaving developers in a permanent migration treadmill.

The iconic paperclip assistant from Microsoft Office that asked 'It looks like you're writing a letter. Would you like help?' Universally loathed by users but deeply loved in hindsight. Killed after a decade of annoyance, Clippy became the original AI assistant meme and the spiritual ancestor of every chatbot that followed.