REVIEWS / AI MODELS / OWNER INSIGHTS
🦉 WE READ 620 OWNER COMMENTS
GPT-5: what owners actually say
Users see diminishing returns from GPT-5, with soaring power costs, persistent basic reasoning failures, and a growing sense that the LLM scaling wall has arrived
What owners complain about
- Power consumption spiral COMMON
Multiple users flag that newer models like GPT-5 consume even more power than predecessors, with one noting 'a lot of it seems to be just inefficient coding as DeepSeek has shown.' Users feel OpenAI doesn't disclose metrics that would look damning and assume undisclosed numbers are worse than competitors'.
- Basic reasoning still broken COMMON
Users cite persistent failures on trivially simple tasks — the 'how many Bs in blueberry' problem is referenced as an example where the model confidently gives wrong answers. One user says they 'can't even trust ChatGPT to answer a basic question without f***ing up and apologizing,' only to repeat the error.
- Scaling wall reached SOME
Users explicitly state 'GPT-5 has basically proved that we've hit a wall and the belief that LLM will just scale linearly with amount of training data is false.' There's a strong sentiment that Altman's PR push is masking the plateau.
- No true creativity SOME
Stack Exchange users note that 'as far as LLMs go, they cannot be creative... they will always only transform already known content.' Single-pass RAG is described as performing poorly, requiring multi-agent generate-and-verify loops for anything non-trivial.
- Integration friction FEW
Developers report specific version mismatches: Spring AI needs 1.1.0-M1 or later for GPT-5 support, older versions silently fall back to gpt-4o-mini. The gpt-5-nano model doesn't accept 'none' for reasoning effort, only 'minimal,' which is undocumented.
What owners love
- Practical coding utility
One developer describes using Claude (Anthropic) to implement an ad pacing system with a PID controller — 'I have no prior experience with PID controllers' — and successfully scaffolding the project in a single session. Frontier models are valued for accelerating implementation of unfamiliar technical domains.
- Mixture-of-experts architecture
Users discuss MoE positively as a scaling approach that lets models store vast information and specialize — 'modules which are better at math, or modules better at prose, or sports' — which is seen as a meaningful architectural advance.
- Emergent capabilities theory
There's acknowledgment that 'emergent capabilities theory sort of supports' the possibility that sufficiently large transformers could reach general intelligence, keeping some optimism alive about the trajectory even among skeptics.
Surprising patterns
- A psychiatrist offered a detailed theory that AI lacks intelligence because it has no childhood dependency, no experience overcoming suffering, no need for food or restroom breaks, and no concept of loss — framing the gap as developmental rather than computational.
- Users draw a sharp distinction between 'intelligence' and 'agency,' arguing that superintelligence does not inherently imply motivation or desire to exert power — challenging the core assumption of AI doomer arguments.
- Even harsh critics concede that benchmarks comparing model intelligence are hard to design reliably, with one user proposing comparative tests using old vs. new models and computing joint sequence probabilities to validate benchmark trustworthiness.
WHO SHOULD SKIP IT
Buyers who need reliable, correct answers on simple logic and reasoning tasks, who are sensitive to environmental cost, or who expect each new model generation to be a quantum leap rather than an incremental improvement should look elsewhere.
Synthesised from 620 real owner comments across 5 platforms. Every point is grounded in the comments — no marketing, no AI guessing. How we do it →