SUMM

GPT-5 has been a polarizing launch, with reactions ranging from praise to disappointment.
Some users prefer GPT-4.0 and are unhappy about its retirement, valuing its personality and familiarity.
Sam Altman acknowledges the need for model customization for different users.
Simplicity benefits novices, but advanced users want to select models based on use case.
OpenAI plans to focus on stability, then adjust GPT-5 to have a "warmer" personality.

Artificial Analysis ran eight independent evaluations on GPT-5 with early access.
GPT-5 offers multiple "reasoning effort" configurations: high, medium, low, minimal.
These configurations greatly affect intelligence, token usage, speed, and cost.
High reasoning effort uses 82 million tokens, while minimal uses only 3.5 million, making it highly efficient.
GPT-5 scores 68 on the AI index, setting a new standard and topping most benchmarks.
GPT-5 is particularly strong in long-context reasoning, aiding agentic coding and large codebase analysis.
The artificial analysis index ranks GPT-5 high at 69, leading other models and restoring OpenAI to the number one AI spot.

Some graphs shown during the GPT-5 launch were inaccurate, leading to "GraphGate."
Bar heights did not correspond to numbers—became a meme within the community.
Despite this, the model's capabilities are generally acknowledged as strong.

Sponsor segment highlights Chat LLM by Abacus: a unified platform for testing multiple models, including GPT-5.
Features include automatic model routing, document chat, image/video generation, and a multi-LLM agent for various tasks.

LM Arena also ranks GPT-5 as number one for text, web development, vision, coding, math, and creativity tasks.
GPT-5 tested under the name "summit" achieves ELO 1481, leading competitors by 20 points.

Some in the industry argue that saturated benchmarks make further comparisons meaningless.
Theo GG and others emphasize the importance of real-world instruction-following and "vibes."
GPT-5 saturates certain tests like AMI 2025 (scoring 100%).
Some maintain that benchmarks can be updated for new qualities, while others value feel over data.

Opinions differ sharply: some see GPT-5 as a breakthrough, others call it a flop or worse than alternatives.
Stage Hand's API tests show GPT-5 trailing Opus 4.1 on both speed and accuracy for browsing use.
Open-source models (e.g., GPTO OSS 12B) sometimes perform surprisingly well.

Reviewers praise GPT-5's personality—direct, less sycophantic, better at pushing back.
Latency and hallucinations have improved, but some still prefer Cloud Code+Opus for coding.
Introduction of a model router in GPT-5 dynamically selects the best sub-model for each prompt.
Manual override options allow for faster answers if needed.

Jailbreak attempts on GPT-5 still succeed with some effort and tricks, demonstrating persistent vulnerabilities.
Customization has increased (e.g., chat color options), but some see this as a shift toward mainstream appeal over innovation.

XAI's Tony Woo claims Gro 4 leads on certain benchmarks like ARC AGI.
OpenAI seen as a respectful competitor, but other players emphasize smaller team accomplishments and continual rapid model releases.

Pricing for GPT-5 is significantly lower than competitors: $1.25/million input tokens, $10/million output.
Lower price increases adoption and ecosystem growth; Opus 4.1 is much pricier ($15/$75 per million).
Improved for computer use agents; GPT-5 passes tasks that stymied GPT-4.0.

GPT-5 shines in state-of-the-art evals and is hyped for writing quality and affordability.
Some large code tasks (automatic refactoring) result in non-functional but "beautiful" code, indicating practical limits.
Medical queries are improving, raising concerns and changing doctor-patient dynamics.
Discussions on AGI and the socioeconomic impacts of compute costs—debates about future leverage and societal classes.

The new GPT-5 lineup simplifies naming, mapping old model names to new versions.
Some users and industry experts consider GPT-5 disappointing, preferring Claude 3.5 or feeling progress is stalling.
Critique centers on diminishing returns; future progress may depend on better software scaffolding around raw model power.
Memes and humor abound on industry forums about job security and AI’s advancing capabilities.

On some benchmarks, Gro 4 remains ahead (e.g., ARC AGI), but GPT-5 leads most others.
Ongoing competition among AI labs is viewed as beneficial to users.
Video ends with a call to like and subscribe.

The Industry Reacts to GPT-5 (Confusing...)