晶晶香港分公司
Hk Recruitment Old Post
LAsincerely invites you to join. To provide you with better service, we hope you can spare 1 valuable minute to fill out the questionnaire below, so we can better understand your needs.
Wangjiao Wet Market
58 Daojia sincerely invites you to join. To provide you with better service, we hope you can spare a precious minute to fill out the questionnaire below, so we can better understand your needs.
Author: Duan Xiaocao
Link: https://www.zhihu.com/question/1936601885311239615/answer/1936983438256244569
Source: Zhihu
Copyright belongs to the author. For commercial use, please contact the author for authorization; for non-commercial use, please credit the source.
Review: GPT-5 returns to the top, but it is not the AGI everyone expected. The most important points first:
1. Accessibility: The barrier to cutting-edge intelligence has been lowered again.
All users can access GPT-5, and even free users can experience it to a limited extent (beyond the limit, it will downgrade to GPT-5-mini).
Sam Altman has fulfilled his promise from February, lowering the barrier to experiencing cutting-edge intelligence—this deserves respect. The rumor circulating yesterday that GPT-5 would cost $200 has been confirmed as fake news.
2. Pricing: Lower than previous GPT/o generations, closely competing with Gemini 2.5
Offers three API variants: GPT-5, mini, and nano (gpt-5-pro has no API), priced as follows:
GPT-5: $1.25 per million input tokens, $10 per million output tokens
GPT-5 Mini: $0.25 per million input tokens, $2.00 per million output tokens
GPT-5 Nano: $0.05 per million input tokens, $0.40 per million output tokens
Frankly, this pricing is highly competitive—lower than GPT-4.1/GPT-4o, on par with Gemini 2.5, lower than Claude Sonnet 4, and far below Claude Opus 4. This means that GPT-5, while matching Gemini 2.5 in price, shifts the current model's "Pareto frontier" upward.
3. Performance: Returns to SOTA, but without "epoch-making" advantages
On LMArena, GPT-5 achieved first place in all categories and overall rankings:
GPT-5 reclaimed the top spot from Gemini 2.5 Pro:
However, the lead isn't large enough—Gemini 3! Gemini 3! Gemini 3! Moreover, GPT-5’s core this time is fusion and built-in routing, with internal model relationships as shown in the diagram:
4. Revealing the true identity of previously anonymous models
The "summit" on LMArena is GPT-5
"Horizon" on OpenRouter is part of the GPT-5 family (though it's still unclear whether it's GPT-5 or mini)
Practical testing. Not yet available in ChatGPT, currently conducting side-by-side tests on LMArena (vs Gemini 2.5 Pro).
Generate an SVG of a pelican riding a bicycle
GPT-5 (essentially consistent with the previous Horizon test, arguably the best among current models):
Gemini 2.5 Pro (seems a bit too fat):
Draw a Pikachu playing basketball
GPT-5 (features mostly accurate):
Gemini 2.5 Pro (expressive, but...):
Code generation capability. On https://gpt-examples.com/, there are many official prompts and programming examples:
Take one prompt at random and compare with Claude.
This is GPT-5's example:
This is Claude Sonnet 4's result:
All necessary features are present, though the blue-purple gradient clearly indicates Claude; GPT-5's page elements might be more appealing:
API. In addition to showcasing example prompts and outputs as above, making it easy for users to copy and test directly, OpenAI has specifically written documentation for using GPT-5.
Variants and best use cases:
gpt-5: Complex reasoning, broad world knowledge, and code-intensive or multi-step agent tasks
gpt-5-mini: Cost-optimized inference and chat; balances speed, cost, and capability
gpt-5-nano: High-throughput tasks, especially simple instruction following or classification
New features and parameters in the API:
The reasoning.effort parameter controls the number of reasoning tokens generated by the model before producing a response. Early reasoning models like o3 only supported low, medium, and high:
- low prioritizes response speed and fewer tokens
- high emphasizes more comprehensive reasoning
The new minimal setting generates very few reasoning tokens in scenarios requiring the fastest time-to-first-token.
Compared to generating none, allowing the model to produce minimal tokens when necessary typically results in better performance. Default is medium.
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-5",
input="How much gold would it take to coat the Statue of Liberty in a 1mm layer?",
reasoning={
"effort": "minimal"
}
)
print(response)
ARC-AGI-2 released GPT-5's results, which did not surpass Grok 4's score, ranking second:
Today's live demo had multiple issues: two diagrams were incorrect (the published blog is fine, but the presentation wasn't reviewed—very unprofessional), and one demo case was criticized for violating physical laws.
First error:
Correct diagram:
Second error (left: 50 < 47.4):
Correct diagram:
Third error, regarding the airplane wing lift demo:
In summary: GPT-5's overall capability is indeed currently the strongest, and considering accessibility and cost-effectiveness, it can be said to be the most usable model today, but the extent of its advantage is not overwhelmingly large. However, it lost to Grok 4 on ARC-AGI-2, and its coding ability may not consistently beat Claude 4; writing style also depends heavily on personal user preference (for example, I personally prefer the earlier Grok 3 and current Gemini 2.5 Pro, and haven't used ChatGPT as my primary tool for about three months).
Regardless, GPT-5 has finally been released. I feel somewhat let down—it's not as groundbreaking as imagined, more like a model that integrates GPT-4o and o3, reduces hallucinations, updates the dataset, and catches up to the current leading-tier capabilities. Of course, since it's just launched, we need to use it more, compare more, and observe community feedback to truly determine how good GPT-5 really is.
Edited on 2025-08-08 09:49
HK$5,000-10,000 Per Month with Commission
Favourites
Share
Show Map
