OpenAI has just open sourced the powerful gpt-oss model series, and then launched the much-anticipated GPT-5. The official slogan is still the familiar "smartest, fastest and most useful so far", but its core has undergone qualitative changes. GPT-5 not only brings built-in "thinking" capabilities, but also announces its "onboarding" to everyone, including free users, in an unprecedented open attitude.
As soon as the news came out, the entire technology circle was instantly ignited. Developers and users who have been waiting for a long time shouted "victory", and the strategic departments of major competitors must have spent a sleepless night.
What new heights has GPT-5 achieved this time?
If GPT-4 is like a doctor who is knowledgeable but sometimes a bit rigid, then GPT-5 is more like a team of experts with a PhD mind who have learned flexibility and in-depth thinking. Its progress is not only a linear increase in performance, but also an innovation in architecture and concepts.
It is no longer "one muscle" and has been upgraded to an intelligent dispatching system.
The core architectural change of GPT-5 is the introduction of a “unified intelligent system”. This is no longer a single model, but an intelligent, layered processing center. You can think of it as an efficient project manager who manages two core employees:
“Quick Gunner” (efficient default model):Responsible for quickly responding to and processing most routine requests to ensure a smooth user experience.
“Thinker” (GPT-5 thinking):Specially designed to gnaw hard bones and deal with difficult problems that require in-depth reasoning, complex analysis and creative conception.
The system’s built-in “real-time router” analyzes the intent of user input, the complexity of the problem, and whether a tool needs to be called in real time, and then intelligently decides to whom to assign the task.
Even better, if you feel that the AI's answer is not in-depth enough, you can directly add instructions such as "Please think deeply about this question" in the prompt word to force the "thinker" model to be called for longer and more in-depth calculations.
When usage reaches the upper limit, the system will seamlessly switch to the smaller "GPT-5 mini" model to ensure service continuity.
All-round, performance comprehensively surpasses the previous generation
OpenAI used a bunch of dazzling benchmark test data to unabashedly demonstrate GPT-5’s comprehensive surpass of previous generation models (such as GPT-4o and o3). The "toothpaste" squeezed out this time was indeed a bit too much.
Coders ecstatic?In the SWE-bench test, which measures real-world software engineering capabilities, GPT-5 scored as high as 74.9%, far exceeding GPT-4o's 30.8% and o3's 69.1%. It excels particularly well in front-end generation and debugging of large code bases. In the official demonstration, a web game with complete functions and beautiful interface can be generated in just one sentence, and even the aesthetic details such as code spacing and layout can be better understood.
Science students counterattack:In the AIME mathematics competition test, the accuracy of GPT-5 without using tools reached an astonishing 94.6%, while GPT-4o was only 42.1%. On the more difficult doctoral-level scientific questions (GPQA), GPT-5 also leads by a large margin. This shows that their logical reasoning and abstract thinking abilities have reached new heights.
The problem of "seriously talking nonsense" has changed a bit:The "illusion" problem is a long-standing problem with large models. GPT-5 has made significant progress in this regard. Official data shows that the probability of factual errors in its answers is about 45% lower than that of GPT-4o; after enabling the "thinking" mode, the error rate is about 80% lower than that of the previous generation inference model (o3).
GPT-5 is also more "honest" when dealing with missing information. When it cannot complete a task or lacks key information, it will be more inclined to admit limitations instead of fabricating results out of thin air like previous generations of models.
Don’t be afraid as a worker, AI is not here to replace you
It is better to say "liberation" than "replacement". This time OpenAI clearly demonstrated the potential of GPT-5 as a powerful collaborator. Its goal is not to replace workers, but to liberate us from repetitive and tedious work and become everyone's "super brain."
For developers, GPT-5 is an ideal “AI co-pilot”. It can take over the time-consuming basic coding, debugging and front-end construction work, allowing developers to focus more on more creative core tasks such as system architecture and algorithm optimization. It is not here to steal the rice bowl, but to help make the "rice" bigger and more delicious.
For the majority of professionals, GPT-5 is an empowering tool that allows everyone to become a "super individual." It no longer just passively answers questions, but learns to think proactively.
In the official demonstration, after it analyzes a data report, it will proactively make suggestions for the next step like an experienced data analyst: "Do you want to overlay quarterly goals on the chart?" "Or, can I help you generate a difference analysis dashboard?" "Directly export a one-page PDF report?"
This ability, coupled with the ability to securely connect to applications such as Google Drive and Teams within the company, means that ordinary employees can easily complete data analysis and report production that required professional teams in the past, allowing them to devote more energy to decision-making and innovation.
Pricing and opening strategies to reshape the landscape
If a technological leap is a demonstration of hard power to opponents, then GPT-5’s pricing and opening strategy is a precise attack aimed at reshaping the market structure.
API pricing:The API price of the main model GPT-5 is set at $1.25/million tokens for input and $10.00/million tokens for output. A more economical GPT-5 mini model is also provided.
Fully open:GPT-5 will be the new default model of ChatGPT and will be rolled out to all Plus, Pro, Team and free users. Paying users will have higher usage limits and access to the top-of-the-line GPT-5 Pro model. Free users will automatically switch to GPT-5 mini after reaching the usage limit.
This move of "increasing configurations and reducing prices (for free users)" will undoubtedly bring the already fierce AI war into a "personal combat" stage.
Impact on competitors:Models such as Google's Gemini and Anthropic's Claude will face tremendous direct pressure. They not only have to catch up in terms of performance, but also respond quickly in terms of pricing and product experience. Especially for Claude, which has always been selling long texts and security, its differentiation advantage will be challenged when GPT-5 significantly improves performance and reliability.
How long can the open source banner be carried forward?The open source army represented by Tongyi Qianwen, DeepSeek, etc. has now become the "last hope" for many people. They can continue to hold high the banner of "free and controllable" to attract companies that don't want to be "hijacked" by OpenAI. But the question is, when the performance gap becomes wider and wider, how many followers can this banner attract?
Impact on application layer developers:For application developers building on large models, this is a moment of both opportunity and crisis. On the one hand, a more powerful base model allows them to create unprecedented application experiences. On the other hand, as OpenAI itself increasingly demonstrates its ability to "turn ideas directly into products", the boundaries between basic models and applications are becoming blurred, which will undoubtedly squeeze the living space of some application layer startups.
The release of GPT-5 is not so much another pure performance competition, but rather a signal that the AI arms race has entered a new stage.
When the performance improvement of the model itself gradually reaches the ceiling, how to transform this powerful power into practical applications and build a complete ecosystem has become a must-answer for all players.
In the future, AI competition will no longer just be about running scores, but also a comprehensive competition of application scenarios, product experience and business models.
We may not see too many disruptive "iPhone moments," but we can see more fierce "pixel-level" optimization and "price wars" among major manufacturers.
For us ordinary users, this may be good news: after all, gods fight, mortals can also get benefits by eating melons and watching movies.