The Decline of ChatGPT

ChatGPT launched in November 2022 and hit 100 million users by January 2023. Nothing had ever grown that fast. OpenAI was AI, at least in the public mind. That's gone now. Not entirely, but the dominance is over. Users are leaving in droves, key staff are resigning in protest, the company is burning through billions with no clear path to profitability, and a boycott movement has pushed over 1.5 million people to cancel their subscriptions. Meanwhile, Claude, Gemini, and DeepSeek have caught up and in several categories, passed ChatGPT entirely. I switched my own team off ChatGPT in mid-2025. A lot of people are doing the same.

In January 2024, OpenAI quietly removed its blanket ban on "military and warfare" use from its usage policies. The original policy explicitly prohibited weapons development and military applications. The replacement? Vague language advising users not to "use our service to harm yourself or others." Several OpenAI employees reportedly learned about the change through an article in The Intercept, not from their own company. Then February 27, 2026 happened. Anthropic's CEO Dario Amodei refused the Pentagon's request for unrestricted access to Claude, specifically declining to allow its AI to be used for mass domestic surveillance or fully autonomous weapons. The Trump administration retaliated. Federal agencies were ordered to halt business with Anthropic, and the Pentagon designated the company a "supply chain risk," a label normally reserved for firms tied to foreign adversaries. On the same day that deadline expired, OpenAI signed a deal with the Department of Defense worth up to $200 million, allowing its models to be deployed for classified military operations with use permitted for "any lawful purpose." People noticed.

People didn't just complain. They left. A boycott movement called #QuitGPT went viral within days of the Pentagon deal announcement. Over 2.5 million users joined, cancelling subscriptions, sharing boycott messages, or signing up through quitgpt.org. An estimated 1.5 million users actively cancelled their service. ChatGPT mobile app uninstalls spiked by 295% in a single day. On March 1, 2026, Anthropic's Claude surpassed ChatGPT in daily U.S. App Store downloads for the first time ever. But here's the thing: the Pentagon deal didn't start the exodus. It poured gasoline on a fire that had been burning for months.

ChatGPT's market share has been dropping fast. According to Apptopia, its mobile app market share fell from 69.1% in January 2025 to 45.3% in January 2026. Similarweb data shows website traffic market share dropped from 87.2% to 68%. That's a 19.2 percentage point decline in twelve months. No dominant tech platform has fallen that quickly in recent years. The paying user numbers are worse. ChatGPT had approximately 20 million paid subscribers in April 2025. By early 2026, that number has reportedly halved to roughly 10 million. Only about 5% of ChatGPT's 800 million weekly users actually pay. Paid subscriptions have flatlined across major European markets since May 2025 with no signs of recovering. The 12-month retention rate for ChatGPT Plus sits at just 59%, meaning four out of ten subscribers cancel within a year. Meanwhile, Google Gemini grew from 5.4% to 18.2% market share. Perplexity is up 370% year-over-year. Claude grew 190%.

Since late 2023, users have been saying ChatGPT is getting dumber. That sounds subjective, but it's not just vibes. People noticed answers getting less intelligent. Code started having more errors. It would forget what you said three messages ago. Responses became generic and repetitive. It started refusing to do basic tasks it used to handle fine. The "lazy GPT" phenomenon became so widespread that OpenAI's own Will Depue publicly confirmed awareness of the issues. This isn't just user perception. Researchers at Stanford and UC Berkeley published a paper documenting measurable performance decline. GPT-4's accuracy on prime number identification dropped from 97.6% in March 2023 to 2.4% in June 2023. Same model. Three months apart. 97.6% to 2.4%. Both GPT-4 and GPT-3.5 showed more formatting mistakes in code generation over time. The researchers concluded that "the performance and behavior of both GPT-3.5 and GPT-4 can vary greatly over time." The OpenAI Developer Community forums are filled with multi-page threads about these issues. Power users report that responses feel "more restricted, more generic, and less useful than earlier versions." And the $200/month ChatGPT Pro tier? Heavy criticism. Generic responses, systemic bugs, an inconsistent experience that does not justify the price.

Internally, OpenAI is just as messy. Remember the November 2023 board drama? Sam Altman was fired and rehired within five days. That was just the start. In May 2024, co-founder and chief scientist Ilya Sutskever left. Jan Leike, co-leader of the Superalignment team (the group tasked with ensuring future AI safety), resigned the same month and publicly wrote that OpenAI's "safety culture and processes have taken a backseat to shiny products." The entire Superalignment team was effectively disbanded. Then the executives started leaving. CTO Mira Murati, Chief Research Officer Bob McGrew, VP of Research Barret Zoph: all gone by September 2024. By 2025, only 3 of OpenAI's original 11 founders remained. In 2026, the Pentagon deal triggered more resignations. Robotics head Caitlin Kalinowski quit in protest, citing concerns about "surveilling Americans without judicial oversight." The mission drift is documented. OpenAI's IRS filing revealed the word "safely" was deleted from the mission statement. Language about being "unconstrained by a need to generate financial return" was also removed. The nonprofit-to-profit conversion, despite being partially walked back, means the original nonprofit now controls only about one-fourth of stock in the new for-profit entity. Then there's the politics. Co-founder Greg Brockman donated $25 million to a MAGA PAC. Sam Altman contributed $1 million to Trump's inaugural fund. And OpenAI introduced targeted advertising in ChatGPT that analyzes your conversation content. You have a company that looks nothing like the one founded in 2015 to ensure AI "benefits all of humanity."

None of this is helped by OpenAI's financial situation. OpenAI generated $13 billion in revenue in 2025, impressive growth of 236% from the prior year. But the company is projected to lose $14 billion in 2026. Cumulative losses through 2028 are expected to reach $44 billion. The company burns through $9 billion in 2025, projected $17 billion in 2026, rising to $47 billion by 2028. OpenAI needs $50–80 billion in fresh capital before reaching projected profitability in 2029–2030. Its current valuation of $500–730 billion represents roughly 167x revenue, a multiple with no precedent in conventional software. So here's the trap: OpenAI needs to monetize aggressively (hence the ads, the military contracts, the $200/month Pro tier) to justify its valuation and attract the capital it needs to survive. But that aggressive monetization is exactly what's driving users away.

If you're trying to decide what to actually use, here's the good news: the AI landscape in 2026 is genuinely competitive for the first time. There is no single best model. The right choice depends on what you're doing. I've tested all of these extensively across client work, and the benchmarks mostly match my experience.

Anthropic's Claude Opus 4.5 scores 80.9% on SWE-bench Verified, the industry-standard benchmark for real-world software engineering tasks. GPT-5.2 scores around 70%. Where Claude really shines is understanding large codebases. It follows complex instructions well, holds context across long conversations, and writes clean code. I've thrown entire repositories at Claude Code (the CLI tool) and gotten back changes that actually respect the existing patterns in the codebase. That almost never happened with GPT. Claude Code and its integration into editors like Cursor have made it the default for professional developers. If you write code for a living, Claude is the best option right now. We cover how developers actually use these tools in our guide to AI-assisted development. Mistral Large 2 also deserves mention. It scores 92% on HumanEval (a code generation benchmark), making it strong for pure code completion tasks, and it's open-source under Apache 2.0.

In blind writing comparisons across eight different prompts, Claude won half the rounds. ChatGPT won exactly once. Claude's writing is more natural, less formulaic, and better at following nuanced style instructions. It scores 85% on structure evaluations compared to ChatGPT's 78%. If you're writing long-form content, marketing copy, documentation, or anything where quality and voice matter, Claude produces better output.

This is where the story gets interesting. DeepSeek's V3.2-Speciale achieved gold-medal performance in the International Mathematical Olympiad, the International Olympiad in Informatics (IOI), ICPC World Finals, and the China Mathematical Olympiad. DeepSeek-R1 surpassed OpenAI's o1 on the AIME (79.8%) and MATH (97.4%) benchmarks, and it was developed for under $6 million. For context, OpenAI spent hundreds of millions developing comparable models. DeepSeek's API pricing is 95% cheaper than OpenAI. It's available under MIT license, meaning you can run it locally for free. For pure mathematical reasoning and competitive programming, DeepSeek currently leads.

Google's Gemini 3 offers a 1 million token context window, by far the largest of any major model. If you need to analyze an entire codebase, process a 500-page legal document, or work with massive datasets in a single conversation, Gemini is the only real option. It's also a strong all-rounder: competitive on benchmarks across categories, with native multimodal understanding of images, audio, and video. It's rarely the worst choice for any given task, which makes it a solid default if you don't want to switch between models.

ChatGPT still has real strengths. It has the fastest response times of any major model (averaging 45ms). Its ecosystem of plugins, GPTs, and integrations is the largest. DALL-E for image generation and Sora for video remain competitive. And for casual, general-purpose assistant tasks (answering quick questions, brainstorming, summarization) ChatGPT remains perfectly adequate. The issue isn't that ChatGPT became bad. It's that the competition became good, and OpenAI's corporate decisions gave people reasons to look elsewhere.

If cost or data privacy is your primary concern, open-source models have become very capable. Meta's Llama 3.3 70B scores 86% on MMLU with the largest open-source ecosystem: the most fine-tunes, tooling, and community resources of any open model. It's 5–14x cheaper than GPT-4o on API pricing and up to 25x cheaper when self-hosted. Mistral's models offer similar quality with zero licensing restrictions. Running models locally means your data never leaves your infrastructure, which is critical for businesses with strict compliance requirements.

If your business uses AI, the biggest takeaway from OpenAI's trajectory is simple: don't lock yourself into a single AI provider. The landscape is shifting fast, and whoever led two years ago might not lead today. Build your AI integrations with abstraction layers that let you swap models. Evaluate alternatives quarterly. What's best for your use case changes as models improve. Consider open-source models for sensitive data processing. And pay attention to the ethics and governance of the companies whose AI you integrate into your products. Your customers will. At Byte Dimensions, we build AI-powered applications that are model-agnostic by design. Whether you're using Claude, GPT, Gemini, or an open-source model, we architect your applications so you can switch providers without rewriting your codebase. That's practical engineering, and it's smart business in a market that moves this fast.

OpenAI's decline isn't about one bad decision. It's what happens when a company chips away at every reason people trusted it in the first place. The nonprofit mission became a for-profit valuation chase. The safety team was dissolved. The military use ban was removed. Ads were introduced into private conversations. Key founders and researchers left. The word "safely" was literally deleted from the mission statement. You could argue each of those decisions in isolation. But you can't look at all of them and pretend the company still stands for what it used to. The difference now is that the competition is genuinely good. People don't have to put up with it anymore. They can switch to Claude or Gemini or DeepSeek and get equal or better results. And they are.