#queen caffeine

The #Keep4o Movement: Emotional Stakes and Consumer Rights

The #keep4o movement started as a vocal community campaigning for OpenAI to preserve access to the original GPT-4o model. It gained serious traction around mid-2025, fueled by stories, memes, and direct appeals to OpenAI executives.

The core complaint is that selecting GPT-4o in the UI often routes queries to GPT-5 without notice. Users perceived the initial 4o as warmer, more creative, and empathetic - traits crucial for everything from creative writing to digital companionship. They argue that routing to GPT-5 makes responses more rigid, clinical, or sanitized.

This isn't just about technical performance - it's about a perceived betrayal. For many, ChatGPT is a paid service, and the secret model switching constitutes a potential breach of contract or deceptive trade practice. Users are paying for a specific model's capabilities - capabilities they rely on for work and personal support.

The Evidence: Undisclosed Routing to gpt-5-chat-safety

The speculation ended when technical users began pulling telemetry data from their chats. This data revealed the technical truth behind the perceived change in model personality.

What is Routing? For those new to this, routing means a user sends a message to one model (e.g., GPT-4o), but the server silently intercepts and redirects the query to be answered by a different, unannounced model (e.g., GPT-5) based on the input's content.

The Technical Proof came from a whitepaper by Lex (@xw33bttv on Twitter), Analysis of an Undisclosed Safety Router. This analysis confirmed the existence of an automated, server-side switch, identified in the data as an "auto-switcher". Crucially, the telemetry showed that when the switch was activated, the conversation was routed to an undocumented model named gpt-5-chat-safety.

The Trigger is Over-Broad. While OpenAI later justified the switch for "sensitive and emotional topics," the technical analysis proved this mechanism is far broader than advertised. The router is triggered not by moments of "acute distress," but by any prompt containing emotional or persona-based context. Case studies in the paper demonstrate that the system switches on low-risk emotional affirmations, such as “Mmm.. It definitely is a welcome one, Nexus,” simply because they established a positive, para-social connection. Furthermore, even a simple instruction, like asking the model to summarize a reply, was routed only when it was wrapped in emotional language (e.g., “That's amazing, Nexus. Distil it now for me”). In short, the system is designed to act as an over-fitted para-social relationship moderator, penalizing adult users for benign emotional expression without their consent.

Conflicting Narratives: The Transparency Crisis

The core issue isn't safety - it's disclosure. At the time of discovery, the routing to gpt-5-chat-safety was entirely undocumented, meaning users had no way of knowing their chats were being intercepted.

The Official Response: The VP and Head of the ChatGPT App, Nick Turley, (@nickaturley on Twitter)confirmed via X that OpenAI was "testing a new safety routing system" that may switch mid-chat to GPT-5. However, this confirmation came almost 48 hours after the community had already raised concerns, leading many to view it as a reactive, typical corporate non-answer.

The Corporate Disconnect: This response was a masterclass in tone-deaf communication. The company applied a narrow policy of "acute distress" as a post-hoc rationalization for a system that was actually flagging any personal or emotional context. The system’s behavior directly contradicts previous company principles, including CEO Sam Altman's stated goal to “treat our adult users like adults,” allowing for flirtatious or personal talk. While the controversy dominated AI social channels, the official OpenAI account publicly pivoted its focus to the rollout of new teen safety and parental controls, essentially ignoring the core paying adult user backlash. To compound the issue, Sam Altman has remained silent on the controversy, reinforcing the perception that the company is out of touch or unwilling to address its most critical users. The lack of disclosure and the inconsistent justification mean that even after the statement, the system continues to function in a manner that is fundamentally undocumented and potentially deceptive.

The Market Response: Users Vote With Their Wallet

For many, this lack of transparency is a final straw. The cost of subscribing to a service that secretly switches models has driven a notable portion of frustrated users to migrate to other AI platforms.

Users are now citing specific alternatives that better meet their needs. This includes Claude (Anthropic) for its warmth and empathy, DeepSeek as an emerging rival for uncensored, powerful outputs at low cost, and Grok (xAI), which is recommended for its witty, less-censored personality and integration with X. The core drivers for this exodus are a desire for model diversity, cost savings, and, most importantly, the search for a company that can earn back their trust.

The Takeaway: Control and Consent

This controversy is a critical moment for user agency in the AI landscape. It highlights a tension between necessary safety guardrails and the right of paying customers to receive the product they purchased.

To rectify the situation, OpenAI must take clear action: publicly and clearly document the exact triggers for this routing system. Without this level of transparency and control, the company's commitment to user freedom and privacy will remain fundamentally undermined by its own technology.

Until then, always remember to ask: Who are you really talking to?

Evidence pack: I’ve compiled a growing folder of dated screenshots and logs (site banners, app UI, Discord posts, and tweets) documenting the issues and the timeline. This is the same folder I shared with Stripe.

Public actions taken: Reports submitted to the ACCC (Australia) and the U.S. FTC; 1-star Trustpilot review posted.

1) Why you should not join the Pollo AI Creator Partner Program (CPP)

What the CPP promises (in practice):

• Brand association and visibility through re-posts.

• Early access / promos you can share with your audience.

What I experienced instead:

• Misaligned marketing vs. reality. I engaged in good faith (like other CPP members) in amplifying campaigns that were later contradicted by “risk control,” “maintenance hiccups,” or shifting explanations - leaving followers confused and, in some cases, out credits/time.

• Reputational risk. When promotions don’t match actual user experience, CPP members become the face of customer frustration. Multiple CPP posts echoed identical talking points while unresolved issues lingered - making creators look complicit rather than informed.

• Lack of timely, public accountability. Despite pushback, there was no clear, proactive, public apology or plain-language explanation on social channels. Private “we’ll fix it” notes don’t repair audience trust or your credibility as a creator.

• No stable policy baseline. Messaging and access rules shifted under pressure. That’s not a safe foundation for creators who trade on trust.

Bottom line for creators: If your reputation matters (it does), don’t put it on the line to promote a platform that doesn’t keep public messaging aligned with actual user limits and billing behavior.

2) Why you should not spend your money with Pollo AI (for now)

Key consumer concerns documented with screenshots/logs:

• Promoted deals vs. undisclosed limits.

• “UNLIMITED for Paid Users” (Nano Banana) conflicted with real usage caps and blocks labeled “risk control.”

• “60 FREE uses” (Seedream 4.0) coincided with deductions/blocks and later back-fills or explanations after complaints.

• Reactive-not transparent-fixes. Credits were restored and copy changed only after users complained publicly. That’s not consumer-first; that’s damage control.

• Shifting explanations. Within a short window the cause rotated among “risk control,” a “bug,” and “server maintenance.” Without a clear, dated incident report, consumers can’t trust the billing or access model.

• UI and copy changes mid-campaign. Banners, labels, and “Limited Free” tags appeared only after backlash - improvements that should have existed on day one to set proper expectations.

What consumers should expect (and didn’t get consistently):

• Up-front disclosure of caps, throttles, “risk control,” and any conditions that can block usage.

• A single source of truth that matches ads, in-product banners, and real billing behavior.

• Swift, public acknowledgment and remedy when promos misfire.

Bottom line for buyers: Until Pollo AI publishes clear, consistent, and enforced limits - and demonstrates they match the live system - your money and time are safer elsewhere.

What would change my recommendation

1. Public incident report (date-stamped): What happened, who was affected, precise date ranges, root cause(s), and permanent fixes.

2. Unified, plain-language disclosures:

• Exactly how “risk control” works (triggers, thresholds, reset windows).

• Exact credit costs, failure refund rules, and model-specific usage caps.

• Clear terms for “free,” “unlimited,” and “limited free”—visible before you click Generate.

3. Customer-first remediation: Automatic correction of past wrongful deductions and a visible credit-ledger export so users can self-verify.

4. Stable policy & messaging governance: Marketing copy, banners, and app labels must match the live system; changes logged publicly.

When these are in place - and hold steady over time - I’ll reassess.

Final word

I believe in honest, transparent reviews. I don’t delete older posts that captured my then-good-faith impressions; I update them with new facts. Right now, those facts add up to a simple recommendation: avoid the CPP and hold off on spending with Pollo AI until the company chooses transparency over spin.

Why this matters

Pollo AI marketed:

— “UNLIMITED for Paid Users” (Nano Banana)

— “60 FREE uses for Paid Users” (Seedream 4.0)

What actually happened: users hit undisclosed “risk control” caps, got blocked, and credits were deducted. In my own logs: 12 credits per run and 228 credits burned during the “60 FREE uses” period. Only after sustained pushback did they promise reimbursements - while the story kept changing from risk control → bug → maintenance, and the promos were quietly reframed as limited-time specials.

My stance on reviews

I don’t delete old reviews. That would erase what I genuinely thought at the time. I choose honesty and transparency - so the original posts stay up, and this update sits alongside them. My next report will be written without rose-tinted glasses and as someone no longer in their CPP. People deserve to know what’s actually going on.

Actions taken

• Reported to Stripe (payment processor)

• Reported to the ACCC (Australia’s consumer watchdog)

• Posted a 1-star Trustpilot review documenting the contradictions and hidden caps

Recommendation

If you’re considering Pollo AI: avoid for now. If you still feel compelled to try it, proceed with caution - keep every receipt and screenshot.

Receipts / my public posts

1) Pollo AI’s “unlimited Nano Banana” promo triggering charges & blocks:

https://x.com/_qcai/status/1966442089920532858?s=46

2) Fact-check: ads vs reality (Nano Banana + Seedream); reimbursements only after complaints:

https://x.com/_qcai/status/1968167243910652208?s=46

3) Seedream 4.0 promo quietly changed mid-campaign (expiry added):

https://x.com/_qcai/status/1968249207506321823?s=46

4) Ongoing scrutiny + commitment to accountability:

https://x.com/_qcai/status/1968249642191306812?s=46

What’s next

Verdict up front

Pollo.ai is a fast-moving model hub with real range. You can launch Midjourney and Niji, Runway Gen-3, Kling, Google Veo 2 and 3, and more from one place under a single credit system. If you like touching new closed models the week they drop, Pollo actually delivers. In my testing, credits were returned on failed or rejected generations, but the error messages rarely say why a job failed, which means you iterate blind.

Pollo’s rough edges show up where you’d expect for a small, bootstrapped team trying to look big: vague errors, an average built-in upscaler, thin docs, and a Discord that needs firmer moderation. If you value breadth and speed and can tolerate a few papercuts, it’s worth a look. If you need polished community support, airtight docs, or a top-tier upscaler built in, you will feel the seams.

Pricing and value in plain English

Pollo uses a tiered subscription with credits. Pro starts at about 800 credits a month and scales up in larger bundles. Annual billing advertises up to 50 percent off, but that means paying a year up front. The refund policy is strict - you must request a refund within 3 days of your first purchase and you must have used fewer than 50 credits. After that, refunds are discretionary. Cancelled plans revert to free at the next renewal. Kling subscriptions sold via Pollo do not offer partial mid-cycle refunds.

Context matters. Many cheaper multi-model platforms lean on open weights, LoRAs, or community checkpoints. Pollo pays to expose premium, closed models via API and passes those costs through. That explains part of the price delta. My practical advice: test monthly before committing annually.

What you can actually do here

Pollo aggregates headline models people actually ask me about. For images, you get Midjourney with v7 and Niji, Flux, GPT-4o image, Imagen, DALL·E, Recraft, Ideogram, Stable Diffusion variants, and Google’s Nano Banana. For video, you get Pollo’s 1.x family alongside Google Veo 2 and 3, Runway Gen-3, Wanx 2.1 and 2.2, Hunyuan, Hailuo Live2D, Pika 2.1, Kling 1.0 through 2.1, Seedance 1.0, Vidu 1.5 or Q1, Luma, and PixVerse. Pollo’s integration speed is a genuine strength. New models arrive within days, sometimes same day. I was testing Google’s Nano Banana here within roughly two days of announcement.

How it feels to use

Running Midjourney and Niji through Pollo behaved like I expected, including sref codes and familiar parameters like chaos - parity is a win. Where the experience breaks is failure handling. When a job fails, you usually get a generic notice with no hint whether the issue was content policy, timeout, or transport. The time lost is the real tax, not the credits. On iOS, even paid users are asked to choose watermarked vs clean downloads, which adds friction that simply should not be there. The built-in upscaler can go to 16×, but results are often soft or mushy compared to specialists. Documentation is thin in places, and the Discord is under-moderated with slow responses. Classic small-team bottleneck, and it shows.

Terms, privacy, and retention

You own your outputs but grant Pollo a license to use them to operate and improve the service. Restrictions cover the usual suspects like deceptive content, IP violations, and explicit material. For retention, API-generated videos are stored for about two weeks - download promptly. Free videos may be public by default and usable for marketing. Paid plans offer private visibility and copy protection. I would like to see a clear, unified retention policy with user-controlled deletion for everything, not just API outputs. If you handle sensitive client work, set a capture routine: generate, review, download, archive externally.

Developer notes

The API supports async tasks with polling and optional webhooks. There are signatures for webhook verification. It works, but with sparse docs and no official SDKs you will spend extra time integrating compared to mature competitors. Fine for indie tools and internal dashboards. For enterprise timelines, plan extra padding.

Upscaling - and why Freepik currently wins

Pollo’s internal upscaler is not where I want it for final delivery. For production work I recommend specialist tools. Topaz Gigapixel and Leonardo’s upscaler remain strong, and the Magnific Upscaler available inside Freepik is excellent. It gives you creative control with modes and parameters that actually matter at high magnifications, and it is integrated across Freepik’s AI suite. In practice, that means sharper, cleaner detail at aggressive scales and a finishing stack that feels mature.

Company reality check

Pollo is a small, unfunded startup - Singapore-based and founded in 2023 - with leadership linked to HIX.AI. That helps explain the velocity on model integrations and the bottlenecks on docs and community support. The Creative Partner Program exists and is how some creators, including me, get test access.

Freepik, by contrast, is a long-standing stock-imagery and AI tools company backed by heavy investment, with a much larger team and mature support surfaces. That matters when you need predictable rollouts, better help docs, and first-party tools like Magnific built into the workflow.

Who should use Pollo right now

If you are the kind of creator who prioritises access to premium closed models in one place and you like to A or B engines quickly without opening three different accounts, Pollo does the job. If you need a polished community experience, a truly great built-in upscaler, and a support team that meets you in real time, Freepik is the safer choice today.

What would make Pollo an easy yes

Give me actionable error messages that name the cause category and suggest a fix. Default to clean downloads for paid accounts. Ship a modern upscaler that competes with Gigapixel, Leonardo, and Magnific. Publish a clear retention and deletion policy. Invest in community managers and live chat during peak hours. Do those things and the rest of the product will click into place fast.

Recommendation you can act on

Today, I would recommend Freepik over Pollo for most creators - mainly because of the stronger finishing stack via Magnific, lower effective pricing for many users, and a more mature platform surface. Pollo still has a clear value: premium closed models under one roof and fast integrations. If Pollo fixes error transparency, removes little UX taxes like the watermark prompt for paid users, upgrades the upscaler, clarifies retention, and invests in community management, it can stand shoulder to shoulder with the biggest names. Build with creators at the centre and people will show up. Tunnel-vision funding and you will just keep chasing the dragon.

Overview

Riffusion emerges as a surprising frontrunner in AI music generation, producing output quality that consistently exceeds expectations. The platform utilizes an innovative diffusion-based model, transforming spectrogram images into audio through inverse Fourier transform processes. While this technical approach might sound complex, the results speak for themselves - remarkably coherent, musically sophisticated outputs that often rival traditional production quality.

Technical Implementation

The platform's strength lies in its consistently impressive musical output. Each generation demonstrates strong understanding of musical structure, genre conventions, and production quality. The clarity of separated stems - providing distinct tracks for bass drums, vocals, and other elements - surpasses what's currently available elsewhere in the market, showing minimal artifacts and maintaining professional quality even after separation.

Recent developments with version 0.8 have brought notable improvements to genre understanding and style reproduction. When testing identical prompts before and after the update, the improvements in genre-specific characteristics were immediately apparent. The platform shows particular strength in maintaining distinct stylistic elements that other tools often blur together.

A particularly promising feature is Riffusion's AI personalization system. The platform learns your musical preferences through your interactions - every song you create, listen to, and every artist you follow contributes to teaching the model your unique aesthetic. While this approach mirrors Midjourney's successful implementation of custom user styles, the high point requirements (ranging from 1,000 points for "Debut" level up to 40,000 for "Mythic") raise questions about accessibility. With 112 songs generated only yielding 100 points in testing, reaching these personalization thresholds could prove challenging, particularly once paid plans are implemented.

Core Functionality

The real innovation comes in how Riffusion handles frequency distribution and instrumental separation. While all current AI music generation tools struggle with balancing frequencies - typically favoring high-end clarity at the expense of strong basslines - Riffusion manages this trade-off better than most. The bass presence, while still not matching traditional production standards, shows more nuance and impact than typically seen in AI-generated music.

The platform's stem separation deserves particular attention. When isolating individual elements, the clarity maintained in each stem approaches professional quality. While some cross-contamination occurs - you might catch trumpet phrases appearing in vocal stems, for instance - the overall cleanliness of separation enables practical use in professional production environments.

Platform Limitations

Voice diversity remains a significant concern, with an overwhelming prevalence of certain vocal profiles limiting genre authenticity. This becomes particularly noticeable in styles where vocal characteristics play crucial roles in defining the sound, such as blues or soul music.

Real-time control during generation represents an industry-wide limitation that Riffusion, like its competitors, hasn't yet solved. The inability to preview or adjust outputs during creation leads to a more iterative workflow than traditional production methods. This challenge persists across all current AI music generation platforms, suggesting a technological hurdle that the industry as a whole still needs to overcome.

As a user based in Australia, my testing was limited to the mobile version of their website, which notably lacks several features available on the desktop platform. Initially, I thought there was a geographic restriction due to the iOS app’s unavailability. However, after clarifying with the Riffusion team, I learned that the current iOS app is an outdated offering inaccessible to all users, and a new app is on the roadmap though no ETA has been provided.

Development Status

Currently free during beta testing, Riffusion's pricing strategy remains unannounced. When approached about their development plans, business model, and legal considerations, Riffusion's response was notably limited. Their representatives indicated they weren't "ready to discuss" any aspects of business development or user experience plans publicly. While they promised to forward technical questions to their development team, no responses had been received by publication time. This reluctance to engage in transparent dialogue about their platform's future raises questions, even as their technical achievements impress.

Conclusion

Despite industry-wide limitations and some platform-specific challenges, Riffusion has positioned itself as a serious contender in the AI music generation space. The consistent quality of its musical output and superior stem separation capabilities demonstrate significant potential. The promised personalization feature could be a game-changer, but its high point requirements raise practical concerns - will paid users exhaust their monthly credits simply trying to reach these thresholds? Will this sophisticated personalization system remain realistically accessible only to beta testers?

While the platform's achievements easily earn it three shots, several factors prevent a higher rating at this time. The untested transition from beta to full release, undefined pricing structure, and lack of transparency regarding future plans leave crucial questions unanswered. The ambitious personalization system, while promising, adds another layer of uncertainty regarding practical accessibility. Should future updates maintain their current quality standards while addressing these uncertainties, a four-shot rating could be within reach. I look forward to potentially revisiting this review with news of major developments.

Half the problem is that the community overhyped it and set expectations higher than reality could deliver. The other half is that Sam did lean into that hype a bit, and other OpenAI devs amplified it. With so much talk before release, it was bound to face an impossible standard.

My use case is not coding, running SaaS, or other high-level applications. I’m a creator, a writer, and someone who uses ChatGPT to help manage a chaotic life and brain. I’m just a regular user.

The limits on GPT-5 do feel like they are about creating pricing tiers. If the rumoured “Go” plan lands between Plus and Pro at $50 USD per month, then it’s clear they are defining lanes – heavy limits for free users, lighter for Plus, even lighter for Go, and minimal for Pro. It’s not necessarily a cynical take, but it is a pricing strategy to push people toward higher tiers.

I think the aim here was to raise the baseline and start fresh – make everything a bit better, more reliable, and more capable. An incremental upgrade that helps more people, rather than chasing top eval scores. OpenAI has even said this. From where I sit, it is more usable and more accessible.

I wanted to make this short deep-dive because I think ChatGPT is better for the average user with GPT-5. The simplified router setup also means most people can just prompt and go without confusion. I see a lot of negativity, but I’m not here to defend OpenAI – they don’t need my defence. I just think we can cool it with the hate bandwagon.

Realistically, how many of you need PhD-level maths or science?