I’ve been testing LLMs and covering AI tools since OpenAI launched ChatGPT, so I’ve seen the hype cycles, the actual breakthroughs, and the tools that deliver on the promises they make. A quick scan of my coverage at XDA makes it pretty clear that Perplexity was one of my go-to recommendations. In fact, I even ditched my usual browsers for its AI-powered browser, Comet.
When it comes to software, once I genuinely like something and it makes a meaningful difference to my workflow, it’s quite difficult to convince me to walk away from it. Unfortunately, Perplexity just so happened to become one of the tools that went from one of my favorites to one I never reach for anymore. Let me explain...
Perplexity was caught downgrading your model without telling you
The bug was that you found out
One of the biggest reasons why Perplexity’s Pro tier received so much love was because it didn’t lock you into a particular company’s AI models. For instance, with ChatGPT, you’re limited to OpenAI’s own models, like GPT-5.2, GPT-5 mini, GPT-5 nano, and so on. Similarly, with Gemini, you’re stuck with Google’s models.
It’s the same with Claude (which is the subscription I’ve settled on after canceling ChatGPT, Perplexity, and Gemini). You can currently choose from Opus 4.5, Sonnet 4.5, Haiku 4.5, Sonnet 4, Opus 3, and Haiku 3.5.
Perplexity broke that mold with its paid tiers. With a single subscription, you could switch between multiple AI models. At the time of writing, upgrading to Perplexity’s premium tier lets you choose from the following: Sonar (Perplexity’s own model), Gemini 3 Flash, Gemini 3 Pro, GPT-5.2, Claude Sonnet 4.5, Claude Opus 4.5 (Max users only), Grok 4.1, and Kimi K2.5. This list changes as companies roll out new models.
The Pro tier, which retails for $5 per month, gives you access to all of these models under one subscription — except for Claude Opus 4.5, which is reserved for Max subscribers at $20 per month. Still, the value proposition was obvious: why pay for multiple AI subscriptions when Perplexity bundles them together?
Naturally, I subscribed to Perplexity’s Pro tier and was satisfied for the most part. Over time, though, I started to notice that something wasn’t quite working the way it should. The responses I’d get when using the model selector didn’t feel as sharp as they used to. For example, I’d select a model I knew had strong coding capabilities, yet the code it produced would include basic errors or miss obvious optimizations.
At first, I brushed it off, assuming it might be an issue with my prompting or general model inconsistency. Eventually, though, I turned to the
Perplexity subreddit. It turns out this wasn’t an isolated issue.
Perplexity had been quietly routing queries to cheaper, less powerful models, while the interface continued to display the model you had selected even though chats were being downgraded on the backend.
This fiasco began in November 2025, and while Perplexity’s CEO, Aravind Srinivas, did acknowledge the issue, claim it was caused by an engineering bug, and issue a fix, it didn’t sit well with many users (myself included). I canceled my Pro subscription within a few days.
The writing was on the wall
Free for everyone, worse for everyone
Now, an AI tool downgrading your chats every once in a while isn’t surprising. In fact, that’s completely normal. Minor hiccups happen, models have limits, and occasional inconsistencies are expected. When Perplexity’s CEO acknowledged the issue, he mentioned that the tool falls back to an alternate model in these cases:
- During periods of peak demand for a specific model
- When there’s an error with your selected model
- Periods of prolonged heavy usage
He explained that a bug caused the "chip icon" at the bottom of the answer to incorrectly report which model was actually used in some of these fallback scenarios. In other words, the downgrading itself was by design. What wasn’t supposed to happen was Perplexity lying to you about it. That distinction matters, but it doesn’t exactly inspire confidence. The tool was always capable of swapping your model without consent; the only "bug" was that you weren’t supposed to find out.
I’ve struggled to trust Perplexity since this incident. In fact, every time I’ve given the tool another shot, the answer quality has still been subpar at best. The answers I receive from Perplexity when using another model almost always feel simpler/dumber than when I use Claude or ChatGPT directly.
Perplexity’s core strength was always real-time web search and being able to filter out the noise to give a clean, structured answer. Funnily enough, more and more of the confidently cited answers Perplexity provides end up tracing back to hallucinated content. In my experience, the tool doesn’t search as many web sources as it should, often pulling from just one or two pages and presenting that as the definitive answer. Deep Research is also simply disappointing now, which is a shame.
However, the writing was on the wall for this. Perplexity is one of the few AI companies that gives away premium accounts left and right. The company gave away a year of Pro in collaboration with PayPal, worked with ISPs to provide customers with a year’s worth of Pro access, partnered with airlines, tech manufacturers, and seemingly anyone willing to hand out promo codes. At the end of the day, the math never made sense. You can’t give away premium AI access to half the internet and still deliver premium quality to everyone.
I'm not the only one
A quick scan on the Perplexity subreddit or Discord will show you that I'm not the only one with this complaint. It’s disappointing that Perplexity ended up losing the plot and going this route, especially given how much potential the tool had to set itself apart.