Why Gemini Pro Subscribers Are Collectively Disappointed: A Deep Dive into 5 Core Complaints

A Paying User's Disappointed Departure

Recently, a long-time Gemini AI Pro subscriber posted a scathing thread on Reddit titled "Gemini cae en picado" (Gemini is in freefall). This user pays nearly €250 per year for a subscription, expecting a premium AI experience — but instead encountered a series of frustrating limitations and quality issues, ultimately choosing to cancel.

What makes this post worth examining isn't that it's a pure emotional vent, but that it systematically identifies several recurring pain points in paid AI products: model downgrading, feature limitations, restricted creative capabilities, and an eroding sense of value. For any AI company trying to retain paying customers, this feedback carries significant weight.

Core Complaint: Paying for Pro, Getting a Downgraded Model

The first and most fundamental issue the user raised is automatic model downgrading. After just "a handful" of queries in Gemini Pro, the system automatically switches him to the much weaker Flash Lite model.

The experience gap from this downgrade is substantial. After being switched to the lightweight model, the user noted:

• A significant increase in hallucinations
• Shallower, less substantive responses
• A dramatic drop in the model's ability to follow complex instructions

"If I'm paying for the Pro model, I expect to be able to use it reasonably — not spend most of my time on a far more limited model." This statement captures the core expectation of any paying user.

Silent throttling and model downgrading are hidden pitfalls common across many AI subscription services. To understand the root cause, you need to appreciate the structural cost tensions of large language model inference. For a flagship model at GPT-4 level, a single inference can consume tens of times more compute than a lightweight model. The cost gap between Gemini Ultra/Pro and Flash Lite is similarly vast — the Flash series was specifically designed by Google as a "distilled" model, using knowledge distillation to compress the capabilities of larger models into smaller parameter counts, dramatically reducing per-inference costs. When providers silently downgrade their most active users during peak load, they are effectively passing infrastructure costs onto the people who need high-performance models the most. Without transparency, this practice is a fast way to destroy paying users' trust — they're billed for the top tier but unknowingly running on a lower tier the whole time.

Gems Features and Instruction-Following: A Double Failure

Beyond the model itself, the user also sharply criticized Gemini's Gems (custom assistant) feature. He stated that instructions set within Gems "work very poorly," with the model frequently failing to follow the custom rules he'd configured.

Gems is Google's answer to OpenAI's GPTs and Anthropic's Projects — a way to create custom AI assistants by presetting a model's role, behavioral rules, and knowledge context through a System Prompt, so users don't have to re-explain context every conversation. Instruction-following capability is one of the key metrics for measuring a language model's practical utility. When a model repeatedly "forgets" or ignores user-defined rules within a Gem, this reflects underlying deficiencies in long-context maintenance and instruction priority handling — issues directly tied to training quality and context window management. The performance gap between flagship and lightweight models on this dimension is especially pronounced, and silent downgrading to Flash Lite makes the problem even worse.

What the user found particularly absurd is that Gems cannot be edited on mobile. He used the word "subrealista" (surreal, ridiculous) to describe this design flaw. For an AI assistant positioned around mobile-first experiences, being unable to edit your own custom assistants on your phone is a glaring inconsistency.

This type of issue reflects a deeper product logic dilemma: when a provider rolls out features quickly but cuts corners on cross-platform consistency and instruction reliability, users don't feel "feature-rich" — they feel "constantly blocked."

Creative Capabilities: The Most Disappointing Part of the Experience

In his post, the user identified creative content generation as the weakest link in the overall experience.

Image Generation: Overly Restrictive, Inconsistent Quality, Forced Watermarks

On human image generation, the user complained that content restrictions are excessively conservative: "For completely legal, harmless requests, human image generation is completely blocked." This points to the well-known problem of over-tightened AI safety policies. Most providers' content classifiers operate on keyword matching, semantic similarity, or pretrained rules — rather than genuinely understanding user intent. To mitigate regulatory risk and reputational damage, providers tend to set rejection thresholds extremely conservatively, causing large numbers of legitimate requests — artistic creation, historical research, fiction writing — to be incorrectly blocked. OpenAI, Stability AI, and others have all faced user attrition from over-restriction and have repeatedly had to recalibrate the boundary between safety policy and usability. This "better safe than sorry" approach is especially costly for paying users — they've paid for advanced features, only to hit walls at the most basic use cases.

Even when images are successfully generated, quality leaves much to be desired. Gemini's image generation is powered by Google's in-house Imagen model series. The user listed the following specific flaws:

• Poor Z-axis (depth) rendering, resulting in flat, two-dimensional images
• Inconsistent overall quality with poor stability
• Lack of consistency across multiple generations
• All generated images carry a prominent Gemini watermark

The forced watermark reflects both commercial copyright considerations and Google's active support for the C2PA (Coalition for Content Provenance and Authenticity) standard — a framework that aims to mark AI-generated content via metadata to combat the spread of deepfakes. However, for paying creative users, forced visible watermarks are fundamentally incompatible with professional output needs. Leading image generation services like Midjourney and DALL-E have already removed mandatory watermarks for paid tiers, making Gemini's approach look especially dated and further eroding users' sense of value. "This is absolutely the limit!" For a paying user who receives a watermarked final product, the practical value — whether for professional or personal creative use — is significantly diminished.

Video Generation: 3 Generations Per Day Is Hard to Accept

The video generation limits are equally frustrating. The user noted that only 3 video generations are allowed per day — and even if he only uses the feature once a month, unused quota does not roll over, and generated results still contain consistency errors.

Gemini's video generation is powered by Google DeepMind's Veo model. Video generation is one of the most computationally expensive AI tasks — generating a few seconds of high-quality video may require as much GPU compute as thousands of text queries, which is why strict quotas exist across all platforms. However, looking at the competitive landscape, dedicated video generation platforms like Runway, Sora, and Kling typically offer more flexible per-minute billing or rollover quota systems. Setting the Gemini Pro video quota at 3 per day with no rollover is genuinely conservative by comparison. "For a subscription at this price point, these limits are hard to justify." This assessment hits the mark: once the subscription price is set high, user expectations for quota and quality naturally rise to match.

From One Case to an Industry-Wide Trust Crisis

What this user feared most wasn't the absence of any single specific feature — it was a general sense that things are sliding downward: "the service is getting progressively worse." He made clear that nearly €250/year was meant to pay for a premium experience, but what greeted him instead was an ever-growing list of restrictions and real, tangible declines in quality. Cancellation became his final answer.

To be clear, this is a single user's feedback on Reddit, and some descriptions reflect personal impressions that may not represent official product specifications or the experience of all users. But the structural contradictions it reveals are worth serious reflection across the industry:

1. Transparency: When model downgrading happens without proactively notifying users, it easily creates a "bait-and-switch" feeling that undermines the foundation of paying trust. Inference cost pressures are real, but how to control costs while staying honest with users is a product design challenge AI providers urgently need to solve.
2. Perceived value: When quota restrictions, forced watermarks, and content moderation pile up, the "premium feel" of a high-priced subscription quickly evaporates. Paying users operate with a completely different mental accounting than free users — their tolerance for every restriction is much lower.
3. Balancing safety and usability: Overly conservative content moderation policies routinely catch large numbers of entirely legitimate use cases in their net. Building classifiers that can more accurately understand genuine user intent while still mitigating extreme risks remains an ongoing technical challenge for content safety teams.

For AI providers, the key to retaining paying users has never been just about topping benchmarks. It's about making users feel, in their real daily use, that every dollar they spend is worth it. When a once-loyal paying user chooses to leave, that itself is the most important signal a product can receive.