全面了解 GPT-5 及其 2025 年最重要的特性
- Aisha Washington
- 6月7日
- 讀畢需時 13 分鐘
GPT-5 是 OpenAI 推出的生成式预训练 Transformer 系列最新版本。它在生成式 AI 模型技术上迈出了重要一步。OpenAI 打造了这一大型语言模型，使其运行更快、更准确、更高效。该模型现在能够回答难题并更好地理解上下文。它还能处理多种输入类型，如文本、图像和音频。下表展示了 GPT-5 在多方面优于旧版 GPT 模型：
关键进展 | 描述 |
速度与效率 | 快 30-50% 得益于更优的神经架构，在处理难题时表现更出色 |
上下文理解 | 更好地理解棘手问题并保持思路连贯 |
准确性与精确度 | 减少错误，更容易区分事实 |
多模态智能 | 可同时使用文本、图像、音频和视频 |
高级推理 | 以更强的思维能力解决难题 |
OpenAI 的 GPT-5 模型采用混合架构，拥有 超过 5000 亿参数。它具备更强的安全特性，可用于更多专业任务。GPT-5 的发布是大型语言模型领域的重要时刻。该模型能更好地处理信息，并为多个领域带来新功能。随着 OpenAI 持续推进 AI 发展，GPT-5 模型为生成式预训练 Transformer 技术树立了新标杆。
要点总结
GPT-5 功能
高级推理
GPT-5 非常擅长解决问题。它能回答数学、科学和人文领域的难题。该模型针对复杂任务进行快速且深入的思考。Amgen 表示，GPT-5 在处理大量细节的工作中 更准确、更快速。Microsoft 在 Copilot 中使用 GPT-5 帮助用户回答难题并进行更长时间的对话。该模型可协助解答健康问题，并充当智能助手。GPT-5 在实际工作中优于旧版模型。
基准/测试 | 描述 | GPT-5 表现 | 对比亮点 |
Humanity’s Last Exam | 横跨数学、人文和科学的难题测试 | 42%（GPT-5 Pro 带工具） | 略低于 Grok 4 Heavy（44.4%） |
GPQA Diamond | 博士级科学问题 | 89.4%（首次尝试） | 优于 Claude Opus 4.1（80.9%）和 Grok 4 Heavy（88.9%） |
HealthBench Hard Hallucinations | 医疗准确性与幻觉率 | 幻觉率 1.6% | 远低于 GPT-4o（12.9%）和 o3（15.8%） |
ChatGPT Prompt Hallucination | ChatGPT 提示中的错误信息率 | 幻觉率 4.8% | 显著低于 o3（22%）和 GPT-4o（20.6%） |
Tau-bench | 模拟在线任务的 AI 代理能力 | 63.5%（航空网站），81.1%（零售网站） | 在部分项目中略逊于 o3 和 Claude Opus 4.1 |
SWE-bench Verified | 来自 GitHub 的真实世界编码任务 | 74.9%（首次尝试） | 略优于 Claude Opus 4.1（74.5%）和 Gemini 2.5 Pro（59.6%） |
GPT-5 在编码、思考和健康回答方面表现出色。它比旧版模型错误更少、更可靠。
多模态输入
GPT-5 可以同时使用 文本、图像、音频和视频。该模型在一次会话中融合文字、图片和声音。这帮助 GPT-5 理解科学图表、收听音频，并根据提示制作短视频。在医疗领域，它可查看患者病史、化验结果和图片以协助医生。在教育领域，GPT-5 利用多种媒体提供个性化课程，帮助学生完成课程。在研究领域，该模型读取不同数据并帮助科学家加快工作。
GPT-5 可以在一次对话中使用文本、图像、音频和视频。
该模型能够 处理庞大且复杂的输入。
人们通过文字、图片、语音和视频与 GPT-5 交流。
GPT-5 支持 OpenAI 工具如用于视频的 Sora 和用于语音的 Whisper。
这些功能使 GPT-5 优于仅支持文本的模型。
GPT-5 的新技能使其在多类型输入使用方面表现最佳。
代理能力
GPT-5 可以像数字工作者一样行动。该模型 在极少帮助下规划并完成任务。它 记住过去的对话和项目细节。GPT-5 代理可使用其他服务、搜索网络、运行代码，并连接 Gmail 和 GitHub 等 API。这些代理可协助客户服务、健康和金融领域的工作。该模型可将任务分配给专业代理、按顺序使用工具，并用文字、图片和视频制作内容。安全规则要求在重大操作前获得许可，并监控不良行为。
描述 | 生产力/工作流影响 | |
GPT-5 可根据高层级目标自主创建并执行复杂工作流，无需逐步指令。 | 减少人工干预，自动化复杂任务，使用户专注于战略决策。 | |
任务委托给专业代理 | GPT-5 可将子任务分配给特定领域代理（如编码、市场分析）。 | 提高精确度和集成度，通过自主利用专业知识简化工作流。 |
持久长期记忆 | 在较长时间内保持上下文，记住过去的交互和项目细节。 | 实现可靠的项目管理、无缝连续性，并减少管理开销。 |
自主执行一系列工具调用并与 API 交互。 | 支持软件工程和业务流程的自动化，加速任务完成。 | |
多模态内容生成 | 按需创建文本、图像和视频内容。 | 加快营销和内容创作工作流，减少时间和成本。 |
GPT-5 的代理技能帮助人们完成更多工作并节省时间。该模型协助大型项目并减少重复任务。
上下文窗口
GPT-5 可以在一次对话中记住大量内容。API 支持最多 400,000 个 token，远超以往。在 ChatGPT 中，免费用户可使用 8K token，Plus 用户可使用 32K，Pro 用户可使用 128K。较大的上下文窗口有助于保持对话清晰和连贯。研究表明，更长的上下文使对话更优质、更连贯。总结功能进一步增强效果。
模型变体 | F1 分数 | 连贯性（%） | 一致性（%） |
longlora-8k | 14.02 | 42.44 | 62.04 |
longlora-8k + 总结 | 15.77 | 53.41 | 68.60 |
ChatGPT（基础版） | 19.41 | 75.00 | 75.48 |
ChatGPT-16k | 19.92 | 78.60 | 79.20 |
ChatGPT-16k + 总结 | 19.29 | 90.04 | 92.44 |
GPT-4o | 20.35 | 87.70 | 82.00 |
GPT-4o + 总结 | 21.02 | 91.12 | 93.29 |
GPT-5 的强大记忆功能有助于长对话和复杂任务。该模型记住细节并保持思路连贯，使其更易于使用。
减少幻觉
GPT-5 在回答中减少错误。该模型使用 特殊训练和方法来降低错误。研究表明这些方法有助于防止错误，尤其在事实关键的场景中。GPT-5 的错误率远低于旧版模型。例如，在医疗领域，其错误率为 1.6%，而 GPT-4o 为 12.9%。在 ChatGPT 提示中，错误率为 4.8%，远优于以往。
GPT-5 比 GPT-4 和其他大型模型更准确。它在数学和科学测试中获得近乎满分。其新技能使其更值得信赖。人们看到更少的错误和更清晰的界限，因此 GPT-5 是重要工作的首选。
GPT-5 的低错误率和更高准确性帮助人们信任其回答。该模型的新技能使其成为迄今为止最可靠的 AI。
GPT-5 的关键特性
GPT-5 因其新功能而与众不同。这些功能帮助人们日常更多地使用该模型。它们使任务、编码和创意工作更轻松。该模型添加的工具改变了人们使用 AI 的方式，并帮助他们完成工作。
Token 限制
GPT-5 比旧版模型能记住更多内容。它可以在一次对话中保留最多 256,000 个 token。这让人们可以进行更长时间的对话或处理大型项目而不会遗忘细节。开发者可针对特殊任务使用最多 1,000,000 个 token。更大的 token 限制帮助模型记住更多内容并保持对话连贯。
功能 | GPT-5 数值 | 对比/备注 |
上下文窗口 | 256,000 个 token | 从 o3 模型的 200,000 个 token 增加 |
更大上下文选项 | 1,000,000 个 token | 在 GPT-4.1 中提供更大容量 |
输入 Token 定价 | 提供 90% 缓存折扣 | |
输出 Token 定价 | 每百万 $10 | 与 GPT-4.1 和 o3 定价相当 |
Mini 层定价 | 每百万 $0.25 | 更经济的选项 |
Nano 层定价 | 每百万 $0.05 | 最具性价比、能力最低的层级 |
GPT-5 即使在大量 token 下也能良好运行。它在最大上下文中 在难题测试中正确率超过 86%。这意味着人们可以将整个项目保留在一次对话中而不会丢失细节。价格也帮助人们更长时间地使用 GPT-5 完成更难的任务。
不可见推理 token 是该模型的另一亮点。这些 token 通过跳过不必要的步骤帮助 GPT-5 更快地思考。该模型减少停顿和大声思考，因此更快且输出更少。这适用于文本、图片和视频。它还能使回答更快但仍保持正确。
该模型跳过额外停顿并专注于重要内容。
这使 GPT-5 在许多任务中更快且更出色。
个性与氛围编码
GPT-5 具备个性化和氛围编码等新功能。这些功能让人们可以改变模型的说话方式。个性功能让人们选择模型听起来友好、严肃或有趣。这使对话感觉更愉快、更真实。
氛围编码是使用 GPT-5 的新方式。人们可以用轻松的方式告诉模型他们的目标和想法。该模型帮助人们来回交流并修改请求。氛围编码帮助人们快速从想法转向可运行的代码。它融合创意思维和技术工作，使构建应用或解决问题更轻松。
人们可以在一次对话中融合大目标、想法和技术需求。
该模型记住旧对话，因此每一步都建立在前一步之上。
这帮助人们持续推进并信任 AI 处理细节。
Google 和 Microsoft 使用氛围编码工具 帮助更多人用 AI 构建。这些工具使非编程人员和业务团队更容易使用 GPT 功能。
语音模式
语音模式是 GPT-5 最棒的新功能之一。该模型现在的说话方式更像真人。人们可以选择模型的说话方式，如改变语调或速度。还有学习模式可用于学习，帮助学生和语言学习者。
许多人表示与 GPT-5 交谈感觉像与真人对话。该模型使用自然的停顿，并能表现出情感。
Aspect | Positive Feedback | Negative Feedback / Limitations |
Conversational Quality | Feels like talking to a real person; natural pauses, emotional responses, flow | AI still interrupts during natural pauses, breaking flow |
Responsiveness | Sub-second response times, natural conversation pace | Occasional audio artifacts and pitch variations |
Educational Use | Excellent for language learning, interview prep, adapts to learning pace | Hallucination issues with complex questions (33% hallucination rate) |
Accessibility | Game-changing for visually impaired or reading difficulty users | Privacy concerns: voice data used for training unless opted out |
Business Integration | N/A | No integration with business tools or custom data |
Cost and Usage Limits | N/A | Daily usage limits frustrating for heavy users; expensive Pro tier |
Voice mode helps people who have trouble reading or seeing. It also makes learning and practicing languages easier. Some people want better business tools and fewer limits, but most think voice mode is a big improvement.
Coding Improvements
GPT-5 is much better at coding jobs. The model can find mistakes that older models missed. It can run helpers in the background to finish hard jobs, which saves time for coders. The model also makes user interfaces that look good and work well, even from simple instructions.
GPT-5 can build whole apps from just one prompt.
The model lets people call functions in any way and pick how much detail they want.
On coding tests, gpt 5 does much better than old models.
The model is more reliable and can handle bigger, harder projects.
OpenAI says gpt 5 can make websites, apps, and games that look great and work well. Early users say the model often adds helpful things they did not ask for. Some coders see that gpt 5 is more careful and less pushy than other ai, but most agree the changes make it easier to make good digital products.
GPT-5’s main features—big token limits, invisible reasoning tokens, personalities, vibe coding, voice mode, and coding upgrades—make it a strong tool for many people. These features help the model support longer chats, more creative work, and better coding.
OpenAI GPT-5 Comparison
Dataset and Training
OpenAI made gpt-5 with bigger and newer datasets than before. The team got new data from websites, books, and articles. This helps the model learn language and facts better. Gpt-5 is different from gpt-4 because it works as a router model. It can use many tools and skills in one system. These include multi-modal processing and deep reasoning. OpenAI also changed the training cutoff. Now, gpt-5 knows about more recent events and facts. This helps the model give answers that are more accurate and fit the question.
OpenAI’s move to a router model in gpt-5 is a big change from older ways. The model now mixes many skills, so it is more flexible and strong.
Bias Reduction
OpenAI keeps working to make gpt-5 less biased. The company uses feedback from people and experts to find unfair answers. They fix these patterns in the model’s replies. OpenAI tests gpt-5 with many questions to check if it is fair. The team also updates the training data to remove bad or unbalanced content. These steps help gpt-5 give answers that are more fair and respectful. OpenAI wants the model to be safe for everyone.
OpenAI checks and updates training data often.
The team uses human feedback to make answers more fair.
Gpt-5 tries not to repeat harmful stereotypes.
Speed and Coherence
Gpt-5 is faster and gives smoother answers than gpt-4. OpenAI made the model’s design better, so it answers quickly and more correctly. The bigger memory window lets gpt-5 remember more in long chats. This means users get smoother talks and fewer mistakes. People see that gpt-5 can keep track of topics and facts for a long time. This makes it easier to have long talks. OpenAI’s changes help the model do tasks and handle information better.
Model | Processing Speed | Coherence in Long Chats | Memory Window Size |
GPT-4 | Moderate | Good | Up to 32K tokens |
GPT-5 | Faster | Excellent | Up to 1M tokens |
OpenAI’s work on speed and smooth answers makes gpt-5 a good pick for people who want quick and reliable replies.
Access and Pricing
Free and Paid Options
OpenAI lets many people use GPT-5 by giving different plans. The free plan lets you chat as much as you want with normal intelligence. This is good for people who just want to try it out. If you need more, you can pick the Plus or Pro plans. The Plus plan gives smarter answers, quicker replies, and gets you in first. Pro users get all the best tools and can do harder jobs. The table below shows what each plan gives you:
Tier | Description |
Unlimited chat at standard intelligence, best for everyday questions | |
Plus Tier | Higher intelligence, faster replies, and priority access |
Pro Tier | Full access to GPT-5’s highest intelligence and advanced features for professionals |
People who pay can use voice mode, multimodal tools, and memory that remembers them. Free users might have to use GPT-5 mini if they ask too many hard questions. These plans help everyone use GPT-5, from kids in school to big work teams.
API and Platforms
Developers can use GPT-5 with OpenAI’s API or Microsoft Azure OpenAI Foundry. These places let you pick from GPT-5, GPT-5 mini, or GPT-5 nano. Each one costs a different amount and does different things. For example, GPT-5 costs $1.25 for a million input tokens and $10 for a million output tokens. Mini and nano are cheaper for smaller jobs. The API now has new tools, lets you control how much it thinks, and change how much it says. Many developers like how well it codes and solves problems. Some companies, like Cursor and Vercel, say it works better and makes fewer mistakes.
Microsoft 365 Copilot Integration
Microsoft 365 Copilot uses GPT-5 to help people work better. Companies get more done and save money. Big companies say they get back 116% of what they spend in three years. Small companies get even more, up to 353% back. People save about 9 hours every month, and 70% feel they do more work. Copilot helps teams catch up on meetings and makes work better. The chart below shows these numbers:
Using GPT-5 in Microsoft 365 Copilot lets more people use smart AI tools and helps companies reach their goals.
Safety and Reliability
Security Measures
GPT-5 is made to keep people safe and protect privacy. OpenAI uses many safety steps to guard user data. The model follows strict rules to keep information private. The table below lists the main ways GPT-5 stays secure:
Security Measure | Description |
GDPR Compliance | GPT for Work follows GDPR rules. It handles personal data fairly and securely. |
Encryption | Data stays encrypted at rest with AES-256 and in transit with TLS 1.2/1.3. |
Third-Party Security Assessments | Independent audits happen every year. Reports are available for review. |
Access Control | Only approved people can access production data. Admin rights get checked every quarter. MFA is required. |
User Identification | Single Sign-On works with Microsoft 365 and Google Workspace for secure logins. |
Vulnerability Management | Devices use antivirus and patch management. Google Cloud Platform also gets regular updates. |
Certifications | The system holds CASA Tier 3 and Microsoft 365 App Compliance certifications. |
These steps help keep everyone’s data safe. OpenAI checks for problems and updates safety tools often. The team works hard so users get safe-completions each time. People can trust their data is safe when using GPT-5.
Reliability Enhancements
OpenAI made GPT-5 more reliable than older models. The system now gives safe-completions with fewer mistakes. The table below shows how GPT-5 does better than GPT-4o:
Error Type | GPT-4o Rate | GPT-5 Rate | Improvement Description |
Hallucination Rate | N/A | GPT-5 makes far fewer mistakes. | |
Deceptive Answers | 4.8% | 2.1% | GPT-5 gives more honest and clear answers. |
Metric/Feature | Description |
Uptime | GPT-5 keeps running 99.95% of the time, even when busy. |
Request Throughput | The system handles over 10,000 requests every second. |
Infrastructure | A global CDN routes requests in under 100 milliseconds. |
Failover Mechanism | Automatic failover keeps the service up during maintenance. |
Monitoring | Real-time dashboards track performance and spot problems fast. |
OpenAI watches for problems all the time. The system checks how fast it answers and how many tokens it uses. Alerts warn the team if something strange happens. Smart caching and batching help keep answers quick and steady. These features help keep everyone safe and make sure answers are good.
GPT-5’s focus on safety and reliability means people can trust it for important jobs. The system works to give safe-completions and keep everyone’s data safe.
GPT-5 is special because it can solve hard problems and use many types of input. It also acts like a smart helper. People notice that gpt-5 can talk for a long time without forgetting things. It makes fewer mistakes and helps with creative ideas. The model’s big memory and better coding tools help lots of users. In 2025, gpt-5 is useful for school, work, and everyday life.
Many experts say to try gpt-5 for learning, projects, and solving problems.You can start by using gpt, joining the community, or trying gpt-5 in different apps and services.
FAQ
What makes GPT-5 different from GPT-4?
GPT-5 uses more data than GPT-4. It can use text, pictures, and sound together. Its thinking is better and faster. It remembers more things and makes fewer mistakes. People see that it gives better answers and talks more smoothly.
Can GPT-5 understand images and audio?
Yes, GPT-5 can use words, pictures, sound, and video at the same time. This means you can get answers from many kinds of information, not just text.
How does GPT-5 help with coding?
GPT-5 can write code and check for mistakes. It can build apps from simple ideas. It also explains code in easy steps. Many people use it to save time and make fewer errors.
Is GPT-5 safe to use for personal data?
OpenAI uses strong safety tools like locks and passwords. GPT-5 follows privacy rules and keeps your data safe. You can trust it with important things.
How much does GPT-5 cost?
Plan | Features | Price (per month) |
Free | Basic chat, limited features | $0 |
Plus | Faster, smarter, priority access | $20 |
Pro | Full features, best performance | $60 |
