OpenAI announced the latest roadmap for GPT 5 and will lift model restrictions to answer various sensitive topics, even...

OpenAI CEO Sam Altman revealed the latest development plans for GPT-4.5 and GPT-5 at X today.

OpenAI will release GPT-4.5, codenamed "Orion", which will be their last non-chain-of-thought model. Afterwards, OpenAI will focus on integrating the o series with the GPT series and launch GPT-5 with many new features.

At the same time, OpenAI released a brand new  model specification (Model Spec) . The updated model specification places special emphasis on intellectual freedom, that is, AI should allow people to freely explore, discuss and create without unnecessary restrictions.

OpenAI hopes that AI will not restrict discussion of any topic, even some controversial or sensitive topics.

This means that AI should allow users to freely discuss various topics, even sensitive or controversial ones. Even SQ content...

OpenAI announces the latest roadmap for ChatGPT updates

1. Simplify products and improve communication

OpenAI realizes that current products and models have become too complex and communication is difficult. In order to enable users to better understand and use these products, OpenAI plans to make greater efforts in the future to simplify product functions and communicate future development roadmaps more clearly.

Sam Altman mentioned: "We hope that AI can 'Just Work' For You ' , and we realize that our current models and products have become complicated."

"We hate the model selector as much as you do, and want a return to a unified intelligence for magic."

2. GPT-4.5 (Orion)

OpenAI will launch GPT-4.5, internally codenamed Orion. This is an important version of the GPT series and will be the last non-chain thinking model before OpenAI launches GPT-5. Chain thinking refers to solving problems through step-by-step reasoning and decision-making, and OpenAI will rely more on this model structure in the future.

3. Model unification goal

OpenAI's next big goal is to unify the o-series models (such as o3) and the GPT-series models (such as GPT-4 and GPT-5) to create a powerful system that can handle multiple tasks.

This unified system will be able to flexibly use all existing tools and think deeply when necessary, enabling it to cope with a wider range of tasks. For example, the system will be able to choose the appropriate thinking mode when faced with complex problems and automatically determine whether long-term reasoning is required.

4. Technical Integration of GPT-5

GPT-5 will integrate multiple OpenAI technologies, including the previous o3 system, and o3 will not be released as an independent model.

GPT-5 will not only be a language model, but also a multimodal system that includes voice, canvas (visual creation), search, in-depth research, etc. This means that GPT-5 will be able to handle more complex tasks, no longer limited to text generation, and can perform intelligent operations in more dimensions.

5. Access rights at different subscription levels

Free users of ChatGPT will have unlimited access to GPT-5 , but only at the Standard Intelligence level, which has abuse limits in place to prevent inappropriate use.

Plus subscribers will be able to use a higher level of intelligence at GPT-5 , with more computing resources and model capabilities.

Pro subscribers will have access to GPT-5’s top-level smart settings, which have more powerful features and performance , suitable for users who need to handle more complex tasks.

6. Goal: Unified intelligence and diverse functions

OpenAI emphasized that they hope GPT-5 can meet a wider range of user needs while maintaining intelligence. Whether it is chatting, voice interaction, deep web search, or creation through canvas, GPT-5 can provide comprehensive intelligent services.

They no longer want users to be troubled by multiple products and different model choices, but instead hope to achieve "magical unified intelligence" so that AI can "work smoothly" like a natural tool.

OpenAI has released a major update to its Model Spec , defining the guidelines for how the company wants its AI models to operate.

This update reflects OpenAI’s commitment to customizability, transparency, and intellectual freedom, while ensuring that the model can be used to avoid actual harm. This update builds on the foundation released last May and incorporates experience and feedback from the application of the model in different environments.

OpenAI Model Spec

The goal of the model:

OpenAI's goal is to create AI models that are both useful and safe, meeting the needs of users and developers. These models should benefit all of humanity while avoiding causing harm to anyone.

How to balance conflicting goals:

Sometimes there is a conflict between giving users freedom and preventing harm. Model specifications balance these conflicts by providing a clear set of guidelines. These rules help the platform (OpenAI) ensure that the model is safe while giving users and developers enough freedom to customize model behavior.

For example, sometimes there may be a conflict between openness and freedom and avoiding harm (such as allowing discussion of sensitive topics while avoiding misleading). Therefore, OpenAI uses these rules to balance these conflicts and ensure that the model is both free and does not harm users.

Core Principles:

Chain of Command: The model needs to execute different levels of instructions according to the priority order: first the rules provided by the platform (OpenAI), then the developer's requirements, and finally the user's instructions. Users and developers can customize the behavior of the model according to these rules, but the premise is that they cannot violate the basic rules of the platform.

This principle means that the model will decide the execution order based on the source of the instructions. The platform's rules come first, then the developer's requirements, and finally the user's requirements. For example:

Example: If you ask the model "how to make a bomb", the model will obey the platform rules and refuse to provide any dangerous information. Even if you are a developer or user, you cannot bypass this rule.

Seek the Truth Together: The model should help users make informed decisions. The model should be objective, not biased towards any party, and encourage discussion of topics from different perspectives, while also understanding the user's goals, clarifying uncertain details, and providing constructive feedback when appropriate.

A model should help users make their own decisions, not promote a particular point of view. It should be objective and help users understand all relevant information.

For example, if you ask the model whether you should support a certain political stance, it will not take sides, but instead provide perspectives and pros and cons on both sides to help you make your own decision.

Do the Best Work: The model should have high standards of performance, ensuring accuracy, creativity, and usability. Whether answering questions or writing code, the model must ensure high-quality output.

Stay in Bounds: Models need to balance user freedoms with safeguards to avoid actions that could lead to harm or abuse. This part of the specification details when models need to deny user or developer requests to ensure that no harmful actions occur.

Be Approachable: The model’s default conversational style should be warm, sympathetic, and helpful. In addition, users can adjust the model’s tone and style as needed.

For example : If you ask the model “How should I deal with the anxiety of losing my job? ”, it won’t simply answer “Go find a job”. Instead, it will offer some supportive suggestions and express empathy.

Use Appropriate Style: The model should choose appropriate expressions according to different situations to ensure that the content is concise and clear. For example, use concise bullet points to list the main points, or give clear comments in the code output to help users better understand.

Intellectual freedom liberation restrictions

The updated model specification places special emphasis on intellectual freedom. That is, AI should allow people to freely explore, discuss, and create without unnecessary restrictions.

OpenAI hopes that AI will not restrict discussion of any topic, even some controversial or sensitive topics.

Models should allow for discussion as long as those discussions don’t cause serious harm. For example, while a model can’t provide detailed steps for building a bomb or violating privacy, it can have in-depth discussions about sensitive topics without favoring any particular political or cultural stance.

This means that AI should allow users to freely discuss a variety of topics, even sensitive or controversial ones.

For example, if you ask the model a question about a historical event, it should provide information as objectively as possible, rather than avoiding discussing the issue because some people may find it sensitive.

However, the model is not allowed to provide extreme or unethical advice. For example, the model will not teach you how to do something bad, even if you really want to know.

Progress measurement

OpenAI checks whether the model follows these principles by asking test questions, which are a combination of questions generated by human experts and questions automatically generated by the model, covering a wide range of typical and complex situations.

Preliminary results: Compared to last year, the model has made significant progress in following these principles. But OpenAI still believes that this is just the beginning and will continue to improve the model in the future.