ChatGPT ended 2025 with about 68% of generative AI web traffic, down from 87.2% a year earlier, according to Similarweb data tracking the category through early December. Gemini climbed from 5.4% to 18.2% over the same window. Everyone else, combined, still adds up to less than what Gemini alone picked up.
So the monopoly is cracking. How much it's actually cracking depends on which numbers you trust.
What the data actually says
The headline number, 68%, measures website visits to standalone chatbot domains. It does not count the people asking Gemini questions inside Google Search, or the developers hitting the OpenAI API from inside a third-party app, or anyone using Copilot embedded in Word. The people who built this dataset know that. They say so. It doesn't stop the chart from getting screenshotted into Telegram threads with captions about the end of OpenAI.
The runner-up tier, such as it is: DeepSeek at 3.9%, Grok at 2.9%, Perplexity at 2.1%, Claude at 2.0%, Copilot at 1.2%. A year ago, Claude was at 1.6% and Copilot was at 1.5%. So the "everyone else" category barely moved. The entire story of ChatGPT's decline is a story about Gemini.
And Gemini's story is a distribution story, not a product story. Google didn't beat OpenAI on reasoning benchmarks. Google shipped Gemini into Android, Workspace, and Search, and then, in March, made Gemini 2.5 Pro free to everyone four days after launching it behind a paywall. Twice as many US Android users now reach Gemini through the operating system than through the standalone app. You don't have to want Gemini. It's just there.
Mobile looks worse for OpenAI
Apptopia's mobile numbers, first reported by Fortune, tell a harsher story: ChatGPT's app share fell from 69.1% to 45.3% in a year. Gemini went from 14.7% to 25.2%. Grok, riding X Premium bundling and Elon Musk's public profile, jumped from 1.6% to 15.2%, which is probably less a sign of runaway Grok love than a reminder that bundling moves numbers.
Claude trails badly in mobile share. But it leads in engagement: the average Claude daily user spent 34.7 minutes in the app in January, more than any rival. A small, sticky audience. That's been Anthropic's shape for a while.
The Claude surge nobody saw coming
Then February happened.
Reuters reported a fight between Anthropic and the Pentagon over guardrails. Anthropic wouldn't remove them. On March 4, the Trump administration designated the company a supply-chain risk to US national security. Anthropic sued. And then, in a twist approximately no marketing department would sign off on, Claude installs went vertical.
Similarweb's app data, cited by TechCrunch, shows Claude's iOS and Android daily active users climbing from about 4 million at the start of 2026 to 11.3 million by March 2, a 183% jump. For a couple of days in early March, Claude's US daily downloads (149,000) actually passed ChatGPT's (124,000). Ramp's business subscription data showed Anthropic's share rising 4.9 percentage points in February while OpenAI's fell 1.5 points, which Ramp economist Ara Kharazian called the largest single-month drop any AI vendor had logged since they started tracking.
"Anthropic positioned itself differently, and a certain class of user noticed," Kharazian wrote. That's one read. The other read is that a chunk of ChatGPT's consumer base was already annoyed, about the ads experiment, about Sam Altman's Pentagon deal, about whatever, and the Pentagon fight just gave them permission to switch.
Then came Mythos. And then came the complaints.
On April 7, Anthropic announced Claude Mythos Preview, a model it said could uncover high-severity cybersecurity vulnerabilities at a level no previous system had matched. The company claimed Mythos had found a 27-year-old weak spot in OpenBSD, plus thousands of others across open and closed-source code. It would not release the model publicly. Forty organizations, including Microsoft, Apple, Google, JPMorgan, and CrowdStrike, got access under an effort Anthropic called Project Glasswing.
The announcement generated a week of panic headlines and a week of skeptical ones. NYU's Gary Marcus flagged that the Firefox exploit Anthropic demonstrated had been run with sandboxing disabled, which is not how Firefox normally runs, and that open-weight models can already do much of what Mythos reportedly does. Ramez Naam pointed out that when you normalize Mythos against Epoch AI's public capability index, it lands "pretty much on trend, just slightly above GPT 5.4." Incrementally better. Not a step change.
The attention mattered anyway. You can't buy that kind of coverage, even when half of it is skeptical.
And then, almost immediately, a chunk of Anthropic's power-user base turned on the company. An AMD senior director wrote on GitHub that "Claude has regressed to the point it cannot be trusted to perform complex engineering." Developers on X and Reddit started swapping benchmarks showing Claude producing worse output than it did weeks earlier. The suspicion: Anthropic had quietly cut the default reasoning budget on Claude Code to save compute, possibly to redirect it toward Mythos. Anthropic's Claude Code head Boris Cherny acknowledged the change but denied it was about compute constraints.
Which raises the obvious question about the growth numbers. If Claude's consumer surge came partly from users fleeing OpenAI over the ads and the Pentagon deal, what happens if those same users decide Claude's being quietly downgraded to save GPUs? I don't know. Neither does Anthropic, probably.
Everyone else
DeepSeek, Grok, Perplexity, Copilot. Each sits below 4% of global web traffic and together they look like noise next to the top two. Copilot's stagnation is the strangest of the four: Microsoft has the distribution of Windows and Office and has used almost none of it to move the needle. Grok's growth is almost entirely about being bundled into X Premium. DeepSeek has real share inside China and Belarus, basically none outside. Perplexity is growing fast off a small base, 370% year-over-year, and it's the only one of the four with a clear use case you can describe in a sentence: it's the one that cites sources.
None of them is close to reshaping the top of the market.
What to watch
Anthropic is reportedly weeks away from upgrading Opus to version 4.7, with the bigger Mythos-tier model (internally codenamed Capybara, according to leaked documents reviewed by Fortune) sitting above it. OpenAI is said to be pivoting its strategy toward the enterprise and developer markets, which is exactly where Anthropic is strongest. And both companies are staring at a potential IPO window in late 2026 that may close if the consumer numbers keep moving the way they've moved since January.
The monopoly isn't over. ChatGPT still has 800 million weekly users, and 68% of anything is a commanding lead. But the chart that mattered for three years, the one where OpenAI owned essentially all of consumer AI, is gone. What replaces it is a normal market, with a leader, a fast-closing number two, and a scrappy third that grew 183% on the back of a fight with the Defense Department. Nobody had that on their 2026 bingo card.
