Lede

OpenAI did not lose because the brains ran out – it lost because it keeps selling a user experience that feels like unpaid labour dressed as magic.

Hermit Off Script

This is why I say OpenAI is cooked – not because it cannot build smart models, but because GPT-5.2 is the sort of baseline it should have launched as GPT-5 in the first place. Instead, we got the reviewer circus doing the usual clapping routine – “wow, it sings” – and I am meant to pretend a slightly improved, cosmetised version is a revelation. It is not. It reads like catch-up. And a smart model should not require me to become a prompt engineer just to get something normal done. The whole point of language models trained on text, audio, whatever, is to understand the need, and if it does not, it should ask or at least infer intention and context without me writing a legal contract in the prompt box. Honestly, GPT-4o felt closer to that vibe, then they destroyed it – not improved it – cheaper, blunter, stupid, like the early days where you have to explain the same thing twice. So now it feels like they restarted from zero with a cheap model, and even when they stack intelligence on top, it still looks stupid because the foundation is “be cheap, be safe, be bland”. And that is why Gemini and Claude never feel like home for most people as well. The tools demand we adapt to them, instead of the other way round.

P.S. I am not on a Pro subscription, so this roast reflects the day to day experience of a normal user. Pro users may have a different experience.

GPT 5.2 Release, Corporate Collapse in 2026, and $1.1M Job Loss | EP #215

One of the best channel for AI reviews recommended by TheModernHermit.blog

What does not make sense

• Calling a point-release a “new era” while the user still has to babysit basic intent and context.
• Selling “professional work” while quietly implying, “pick the right mode, pick the right tool, phrase it just so”.
• Treating singing demos as proof of intelligence, when they are mostly proof of marketing.
• Optimising for cost and then acting shocked when the product feels cheaper.
• Expecting mass adoption while making normal users do prompt gymnastics to get normal outputs.

Sense check / The numbers

1. OpenAI announced GPT-5 on 7 August 2025 and made it the default in ChatGPT, explicitly positioning it as a unified system that routes between faster answers and deeper “thinking”. [OpenAI]
2. OpenAI announced GPT-5.2 on 11 December 2025, with three variants (Instant, Thinking, Pro) and framed it as best for “professional knowledge work” and “long-running agents”. [OpenAI]
3. OpenAI listed GPT-5.2 API pricing at $1.75 per 1M input tokens and $14 per 1M output tokens, while GPT-5.2 Pro is $21 input and $168 output per 1M tokens. [OpenAI]
4. Reuters reported GPT-5.2 launched on 11 December 2025 after an internal “code red” push, framed as a response to Google Gemini 3, and that rollout begins with paid ChatGPT plans. [Reuters]
5. On the “wow, it sings” front, one widely shared example was a September 2024 demo where a user coaxed OpenAI’s voice bot into singing along to a Beatles track. That is a party trick, not a product strategy. [Ars Technica]

The sketch

Scene 1: The Version Bump Blessing
A stage. A giant “5.2” balloon. Confetti.
CEO: “Behold – productivity!”
User (holding a laptop, exhausted): “Cool. Can it understand what I meant?”
Scene 2: Prompt Engineering Apprenticeship
An office labelled “Normal Task”. A bouncer at the door labelled “Perfect Prompt”.
Bouncer: “Dress code: JSON, constraints, tone, context, examples.”
User: “I just wanted a spreadsheet, mate.”
Scene 3: The Singing Canary
A canary in a cage singing. The room is on fire labelled “UX”.
Reviewer: “Listen to the vocals!”
User: “Yes, lovely. Now, about the flames?”

What to watch, not the show

• Incentives that reward demos over durability.
• Pricing pressure that nudges teams toward smaller, safer, blander defaults.
• Leaderboard culture that measures what is easy to score, not what is nice to live with.
• “Mode sprawl” that turns one product into three personalities and a guessing game.
• The quiet shift from “assistant” to “toolbox”, where the user becomes the project manager.

The Hermit take

If the product needs a priesthood of prompt engineers, it is not intelligent – it is fussy.
Real help feels like relief, not a new hobby.

Keep or toss

Keep / Toss
Keep the push toward better long-context work and reliable outputs.
Toss the theatre – if “wow” needs confetti and a singing demo, you have already lost the plot.

Sources

• OpenAI – Introducing GPT-5 – https://openai.com/index/introducing-gpt-5/
• OpenAI – Introducing GPT-5.2 – https://openai.com/index/introducing-gpt-5-2/
• Reuters – OpenAI launches GPT-5.2 after ‘code red’ push – https://www.reuters.com/technology/openai-launches-gpt-52-ai-model-with-improved-capabilities-2025-12-11/
• Ars Technica – Man tricks OpenAI’s voice bot into a Beatles duet – https://arstechnica.com/information-technology/2024/09/man-tricks-openais-voice-bot-into-duet-of-the-beatles-eleanor-rigby/