{"id":1065,"date":"2026-03-07T20:29:04","date_gmt":"2026-03-08T04:29:04","guid":{"rendered":"https:\/\/51ai.website\/?p=1065"},"modified":"2026-03-07T20:45:59","modified_gmt":"2026-03-08T04:45:59","slug":"gpt-5-4","status":"publish","type":"post","link":"https:\/\/51ai.website\/en\/blog\/gpt-5-4\/","title":{"rendered":"GPT-5.4 \u5168\u6574\u7406\uff1a\u975e\u5e38\u597d\u7528\uff0c\u975e\u5e38\u8d35"},"content":{"rendered":"

 <\/span><\/p>\n\n\n\n

\n

<\/span><\/p>\n

\u8fd9\u6b21\u540c\u6b65\u53d1\u5e03\u4e24\u4e2a\u7248\u672c\uff1a<\/span>5.4<\/span><\/code> \u548c <\/span>5.4 Pro<\/span><\/code><\/p>\n

GPT-5.4<\/span><\/strong> \u662f\u4e3b\u529b\u7248\u672c\uff0c\u5728 ChatGPT \u91cc\u4ee5\u300cGPT-5.4 Thinking\u300d\u7684\u540d\u5b57\u51fa\u73b0\uff0c\u9762\u5411 Plus\u3001Team\u3001Pro \u7528\u6237\uff0c\u540c\u6b65\u4e0a\u7ebf API \u548c Codex\uff0c\u6807\u51c6 API \u5b9a\u4ef7 <\/span>$2.50\/M \u8f93\u5165\uff0c$15\/M \u8f93\u51fa<\/span><\/strong><\/p>\n

<\/figure><\/span>
\u5728 ChatGPT \u91cc\uff0c\u8fd9\u4fe9\u6a21\u578b\u662f\u8fd9\u6837\u7684<\/span><\/figcaption><\/figure>\n

GPT-5.4 Pro<\/span><\/strong> \u9488\u5bf9\u6700\u590d\u6742\u7684\u4efb\u52a1\u573a\u666f\uff0c\u4ec5\u9650 ChatGPT Pro \u548c Enterprise \u7528\u6237\uff0cAPI \u5b9a\u4ef7 <\/span>$30\/M \u8f93\u5165\uff0c$180\/M \u8f93\u51fa<\/span><\/strong>\uff0c\u662f\u6807\u51c6\u7248\u7684 12 \u500d\uff0c\u8d35\u7684\u7406\u7531\u540e\u9762\u7ec6\u8bb2\uff0c\u5148\u8bf4\u6a21\u578b\u672c\u8eab<\/span><\/p>\n

GPT-5.4 \u628a\u4e09\u4ef6\u4e8b\u5408\u5e76\u8fdb\u4e86\u540c\u4e00\u4e2a\u6a21\u578b\uff1a<\/span>\u4ee3\u7801\u80fd\u529b<\/span><\/code>\u3001<\/span>\u901a\u7528\u63a8\u7406<\/span><\/code>\uff0c<\/span>\u539f\u751f Computer Use \u80fd\u529b<\/span><\/code><\/p>\n

\u8fd9\u662f OpenAI \u7b2c\u4e00\u6b21\u5728\u901a\u7528\u6a21\u578b\u91cc\u5185\u7f6e Computer Use\uff0c\u4e4b\u524d\u8fd9\u4e2a\u80fd\u529b\u53ea\u5728\u4e13\u95e8\u7248\u672c\u91cc\u6709<\/span><\/p>\n

\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3 <\/span>1M tokens<\/span><\/strong>\uff0c\u6700\u5927\u8f93\u51fa 128K tokens\uff0c\u77e5\u8bc6\u622a\u6b62\u65e5 2025 \u5e74 8 \u6708 31 \u65e5<\/span><\/p>\n

\u5bf9\u6b64\uff0c\u6211\u7ed9\u505a\u4e86\u4e2a\u56fe<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

GPT-5.4 \u5bf9\u6bd4<\/span><\/figcaption>\n<\/figure>\n
\n

\u4e13\u4e1a\u5de5\u4f5c\u80fd\u529b<\/span><\/h2>\n

OpenAI \u6709\u4e00\u4e2a\u53eb GDPval \u7684\u5185\u90e8\u8bc4\u6d4b\uff0c\u8bbe\u8ba1\u903b\u8f91\u662f\uff1a\u8ba9\u6a21\u578b\u505a\u7f8e\u56fd GDP \u524d 9 \u5927\u884c\u4e1a\u91cc\u5b9e\u9645\u5b58\u5728\u7684\u5de5\u4f5c\u4efb\u52a1\uff0c\u8986\u76d6 44 \u79cd\u804c\u4e1a\uff0c\u4efb\u52a1\u5305\u62ec\u9500\u552e\u6f14\u793a\u6587\u7a3f\u3001\u4f1a\u8ba1\u7535\u5b50\u8868\u683c\u3001\u6025\u8bca\u6392\u73ed\u8868\u3001\u5236\u9020\u4e1a\u56fe\u8868\u3001\u77ed\u89c6\u9891\u7b49\u7b49\uff0c\u7136\u540e\u7531\u4eba\u5de5\u8bc4\u4f30\u8005\u5224\u65ad\u6a21\u578b\u8f93\u51fa\u662f\u5426\u80fd\u548c\u884c\u4e1a\u4ece\u4e1a\u8005\u6301\u5e73\u6216\u66f4\u597d<\/span><\/p>\n

<\/figure>

<\/span><\/p>\n

GPT-5.4 \u5728\u8fd9\u4e2a\u8bc4\u6d4b\u4e0a\u5f97\u4e86 <\/span>83.0%<\/span><\/strong>\uff0cGPT-5.2 \u662f 70.9%\uff0cGPT-5.4 Pro \u662f 82.0%<\/span><\/p>\n

\u5728\u7535\u5b50\u8868\u683c\u65b9\u5411\uff0cOpenAI \u7528\u4e86\u4e00\u7ec4\u6a21\u62df\u6295\u884c\u521d\u7ea7\u5206\u6790\u5e08\u65e5\u5e38\u5efa\u6a21\u4efb\u52a1\u7684\u5185\u90e8\u6d4b\u8bd5\uff0cGPT-5.4 \u5f97\u5206 <\/span>87.3%<\/span><\/strong>\uff0cGPT-5.2 \u662f 68.4%\uff0c\u63d0\u5347\u4e86 19 \u4e2a\u767e\u5206\u70b9<\/span><\/p>\n

\u6f14\u793a\u6587\u7a3f\u65b9\u5411\uff0c\u62ff GPT-5.4 \u548c GPT-5.2 \u7684\u8f93\u51fa\u505a\u76f2\u6d4b\uff0c\u4eba\u5de5\u8bc4\u5ba1\u5728 <\/span>68% \u7684\u5bf9\u6bd4\u91cc<\/span><\/strong>\u66f4\u504f\u597d GPT-5.4 \u7684\u7ed3\u679c\uff0c\u4e3b\u8981\u539f\u56e0\u662f\u89c6\u89c9\u66f4\u591a\u6837\uff0c\u56fe\u7247\u751f\u6210\u7528\u5f97\u66f4\u5230\u4f4d<\/span><\/p>\n

\u8fd9\u6b21\u968f\u6a21\u578b\u540c\u6b65\u53d1\u5e03\u4e86 ChatGPT for Excel \u63d2\u4ef6\uff0cCodex \u548c API \u4e5f\u66f4\u65b0\u4e86\u7535\u5b50\u8868\u683c\u548c\u6f14\u793a\u6587\u7a3f\u7684\u6280\u80fd\u5305\uff08Skill\uff09<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

https:\/\/openai.com\/index\/chatgpt-for-excel\/<\/span><\/figcaption>\n<\/figure>\n

\u4e0b\u9762\u4e09\u5f20\u56fe\u662f GPT-5.2 \u548c GPT-5.4 \u5728\u7535\u5b50\u8868\u683c\u3001\u6587\u6863\u3001\u6f14\u793a\u6587\u7a3f\u4e09\u4e2a\u573a\u666f\u7684\u8f93\u51fa\u5bf9\u6bd4\uff1a<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

\u7535\u5b50\u8868\u683c\u8f93\u51fa\u5bf9\u6bd4\uff1aGPT-5.2 vs GPT-5.4<\/span><\/figcaption>\n<\/figure>\n
<\/figure><\/span>

<\/p>\n

\u6587\u6863\u8f93\u51fa\u5bf9\u6bd4\uff1aGPT-5.2 vs GPT-5.4<\/span><\/figcaption>\n<\/figure>\n
<\/figure><\/span>

<\/p>\n

\u6f14\u793a\u6587\u7a3f\u8f93\u51fa\u5bf9\u6bd4\uff1aGPT-5.2 vs GPT-5.4<\/span><\/figcaption>\n<\/figure>\n

\u5e7b\u89c9\u63a7\u5236\u65b9\u9762\uff0cGPT-5.4 \u662f\u76ee\u524d OpenAI \u65d7\u4e0b\u4e8b\u5b9e\u51c6\u786e\u6027\u6700\u9ad8\u7684\u6a21\u578b<\/span><\/p>\n

\u6d4b\u8bd5\u65b9\u6cd5\u662f\u62ff\u4e00\u6279\u7528\u6237\u771f\u5b9e\u4e3e\u62a5\u8fc7\u4e8b\u5b9e\u9519\u8bef\u7684 prompt\uff0c\u5bf9\u6bd4\u4e24\u4e2a\u6a21\u578b\u7684\u8f93\u51fa\uff1a\u5355\u6761\u58f0\u660e\u51fa\u9519\u6982\u7387\u6bd4 GPT-5.2 \u4f4e <\/span>33%<\/span><\/strong>\uff0c\u5b8c\u6574\u56de\u590d\u5305\u542b\u9519\u8bef\u7684\u6982\u7387\u4f4e <\/span>18%<\/span><\/strong><\/p>\n

\n

Computer Use \u548c\u89c6\u89c9\u611f\u77e5<\/span><\/h2>\n

OSWorld-Verified 75.0%\uff0c\u4eba\u7c7b\u57fa\u51c6 72.4%<\/span><\/strong><\/p>\n

<\/figure>

<\/span><\/p>\n

OSWorld \u662f\u4e00\u4e2a\u684c\u9762\u64cd\u4f5c\u8bc4\u6d4b\uff0c\u6d4b\u7684\u662f\u6a21\u578b\u80fd\u4e0d\u80fd\u901a\u8fc7\u622a\u56fe\u6765\u63a7\u5236\u771f\u5b9e\u7684\u684c\u9762\u73af\u5883\uff0c\u5305\u62ec\u9f20\u6807\u70b9\u51fb\u3001\u952e\u76d8\u8f93\u5165\u3001\u8de8\u5e94\u7528\u64cd\u4f5c\u7b49\u7b49\u3002GPT-5.4 \u5f97\u4e86 <\/span>75.0%<\/span><\/strong>\uff0cGPT-5.2 \u662f 47.3%\uff0c\u4eba\u7c7b\u6d4b\u8bd5\u8005\u7684\u57fa\u51c6\u662f 72.4%\u3002GPT-5.4 \u5df2\u7ecf\u8d85\u8fc7\u4e86\u4eba\u7c7b\u6c34\u5e73<\/span><\/p>\n

\u8fd9\u4e2a\u80fd\u529b\u73b0\u5728\u901a\u8fc7\u6807\u51c6 API \u7684 <\/span>computer<\/span><\/code> \u5de5\u5177\u76f4\u63a5\u63d0\u4f9b\uff0c\u5f00\u53d1\u8005\u4e0d\u9700\u8981\u518d\u8def\u7531\u5230\u72ec\u7acb\u6a21\u578b<\/span><\/p>\n

\u6a21\u578b\u652f\u6301\u4e24\u79cd\u64cd\u63a7\u65b9\u5f0f\uff1a\u7528 Playwright \u8fd9\u7c7b\u5e93\u5199\u4ee3\u7801\u6765\u64cd\u4f5c\u6d4f\u89c8\u5668\uff0c\u6216\u8005\u76f4\u63a5\u63a5\u6536\u622a\u56fe\u7136\u540e\u53d1\u51fa\u9f20\u6807\u952e\u76d8\u6307\u4ee4\u3002\u5f00\u53d1\u8005\u53ef\u4ee5\u901a\u8fc7 developer message \u8c03\u6574\u6a21\u578b\u884c\u4e3a\uff0c\u4e5f\u53ef\u4ee5\u914d\u7f6e\u81ea\u5b9a\u4e49\u7684\u786e\u8ba4\u7b56\u7565\u6765\u63a7\u5236\u9ad8\u98ce\u9669\u64cd\u4f5c\u7684\u5ba1\u6279\u95e8\u69db<\/span><\/p>\n

\u8fd9\u4e2a\u65b9\u5411\u548c OpenClaw\uff08\u4e00\u4e2a\u5df2\u7ecf\u80fd\u7a33\u5b9a\u64cd\u63a7\u7535\u8111\u3001\u6267\u884c\u590d\u6742\u4efb\u52a1\u7684 Agent \u6846\u67b6\uff09\u5728\u8def\u7ebf\u4e0a\u662f\u6536\u62e2\u7684\u3002GPT-5.4 \u628a\u540c\u7b49\u91cf\u7ea7\u7684 Computer Use \u80fd\u529b\u505a\u8fdb\u4e86\u901a\u7528 API\uff0c\u964d\u4f4e\u4e86\u96c6\u6210\u95e8\u69db<\/span><\/p>\n

\u4e0b\u9762\u4e24\u4e2a\u89c6\u9891\u662f GPT-5.4 \u505a Computer Use \u4efb\u52a1\u7684\u6f14\u793a\uff0c\u89c6\u9891\u6ca1\u6709\u52a0\u901f\uff1a<\/span><\/p>\n

\u73b0\u5b9e\u8fd9\u4e2a\uff0cAI \u901a\u8fc7\u901a\u8fc7\u5750\u6807\u70b9\u51fb\u6765\u53d1\u90ae\u4ef6\u3001\u6392\u65e5\u5386\uff0c\u5168\u7a0b\u622a\u56fe\u9a71\u52a8<\/span><\/p>\n

<\/span>

<\/p>\n

GPT-5.4 \u64cd\u63a7\u6d4f\u89c8\u5668\u754c\u9762\u5904\u7406\u90ae\u4ef6\u548c\u65e5\u5386<\/span><\/figcaption>\n<\/figure>\n

\u7136\u540e\u662f\u8fd9\u4e2a\uff1a<\/span>\u7528 Playwright \u81ea\u52a8\u5316\uff0c\u628a\u4e00\u6279\u8bb0\u5f55\u4f9d\u6b21\u63d0\u4ea4\u5230\u5341\u4e2a\u8868\u5355\u91cc<\/span><\/p>\n

<\/span>

<\/p>\n

GPT-5.4 \u6279\u91cf\u586b\u5199\u5341\u4e2a web \u8868\u5355<\/span><\/figcaption>\n<\/figure>\n

\u6d4f\u89c8\u5668\u64cd\u4f5c\u65b9\u5411\uff0cWebArena-Verified \u662f\u4e00\u4e2a\u4e13\u95e8\u6d4b\u6d4f\u89c8\u5668\u63a7\u5236\u80fd\u529b\u7684\u8bc4\u6d4b\uff0c\u540c\u65f6\u652f\u6301 DOM \u5206\u6790\u548c\u622a\u56fe\u4e24\u79cd\u4ea4\u4e92\u65b9\u5f0f\uff0cGPT-5.4 \u5f97\u4e86 <\/span>67.3%<\/span><\/strong>\uff0cGPT-5.2 \u662f 65.4%<\/span><\/p>\n

\u89c6\u89c9\u7406\u89e3\u65b9\u9762\uff0cMMMU-Pro \u662f\u4e00\u4e2a\u6d4b\u591a\u6a21\u6001\u7406\u89e3\u548c\u63a8\u7406\u7684\u7efc\u5408\u8bc4\u6d4b\uff0cGPT-5.4 \u5728\u4e0d\u4f7f\u7528\u5916\u90e8\u5de5\u5177\u7684\u60c5\u51b5\u4e0b\u5f97\u4e86 <\/span>81.2%<\/span><\/strong>\uff0cGPT-5.2 \u662f 79.5%<\/span><\/p>\n

<\/figure>

<\/span><\/p>\n

\u6587\u6863\u89e3\u6790\u65b9\u9762\uff0cOmniDocBench \u6d4b\u7684\u662f\u6a21\u578b\u4ece\u56fe\u7247\u6216\u626b\u63cf\u4ef6\u91cc\u8fd8\u539f\u6587\u672c\u7684\u51c6\u786e\u5ea6\uff0c\u7528\u5f52\u4e00\u5316\u7f16\u8f91\u8ddd\u79bb\u8861\u91cf\u8bef\u5dee\uff0c\u6570\u5b57\u8d8a\u5c0f\u8d8a\u597d\u3002GPT-5.4 \u662f <\/span>0.109<\/span><\/strong>\uff0cGPT-5.2 \u662f 0.140<\/span><\/p>\n

<\/figure>

<\/span><\/p>\n

\u56fe\u50cf\u8f93\u5165\u8fd9\u6b21\u65b0\u589e\u4e86 <\/span>original<\/span><\/code> \u7cbe\u5ea6\u7ea7\u522b\uff0c\u652f\u6301\u6700\u9ad8 <\/span>10.24M \u50cf\u7d20<\/span><\/strong>\u6216 6000 \u50cf\u7d20\u8fb9\u957f\u7684\u5168\u5206\u8fa8\u7387\u8f93\u5165\u3002<\/span>high<\/span><\/code> \u7ea7\u522b\u4e5f\u4ece\u539f\u6765\u7684\u4e0a\u9650\u63d0\u5347\u5230 2.56M \u50cf\u7d20\u3002OpenAI \u5185\u6d4b\u53d1\u73b0\u8fd9\u5bf9\u5b9a\u4f4d\u51c6\u786e\u7387\u548c\u70b9\u51fb\u7cbe\u5ea6\u6709\u660e\u663e\u6539\u5584\uff0c\u5bf9\u9700\u8981\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u622a\u56fe\u7684 Computer Use \u573a\u666f\u5e2e\u52a9\u6700\u5927<\/span><\/p>\n

\n

\u4ee3\u7801\u80fd\u529b<\/span><\/h2>\n

GPT-5.4 \u5728\u4ee3\u7801\u4e0a\u7684\u505a\u6cd5\u662f\u628a GPT-5.3-Codex \u7684\u7f16\u7a0b\u80fd\u529b\u7ee7\u627f\u8fdb\u6765\uff0c\u540c\u65f6\u8ba9\u5b83\u5728\u66f4\u957f\u5468\u671f\u7684\u4efb\u52a1\u91cc\u8dd1\u5f97\u66f4\u7a33<\/span><\/p>\n

SWE-Bench Pro \u662f\u76ee\u524d\u4ee3\u7801\u80fd\u529b\u6700\u4e3b\u6d41\u7684\u8bc4\u6d4b\u4e4b\u4e00\uff0c\u6d4b\u7684\u662f\u6a21\u578b\u5904\u7406\u771f\u5b9e GitHub \u95ee\u9898\u7684\u80fd\u529b\uff0cGPT-5.4 \u5f97\u4e86 <\/span>57.7%<\/span><\/strong>\uff0cGPT-5.3-Codex \u662f 56.8%\uff0cGPT-5.2 \u662f 55.6%<\/span><\/p>\n

<\/figure>

<\/span><\/p>\n

Terminal-Bench 2.0 \u6d4b\u7684\u662f\u7ec8\u7aef\u64cd\u4f5c\u80fd\u529b\uff0cGPT-5.4 \u662f <\/span>75.1%<\/span><\/strong>\uff0c\u8fd9\u91cc\u6709\u4e00\u4e2a\u9700\u8981\u8bf4\u7684\u70b9\uff1aGPT-5.3-Codex \u662f 77.3%\uff0cGPT-5.4 \u5728\u8fd9\u4e2a\u5b50\u9879\u4e0a\u5c0f\u5e45\u56de\u9000\u4e86<\/span><\/p>\n

Codex \u91cc\u65b0\u589e\u4e86 <\/span>\/fast<\/span><\/code> \u6a21\u5f0f\uff0ctoken \u751f\u6210\u901f\u5ea6\u6700\u591a\u63d0\u5347 <\/span>1.5 \u500d<\/span><\/strong>\uff0c\u6a21\u578b\u672c\u8eab\u4e0d\u53d8\u3002\u5f00\u53d1\u8005\u901a\u8fc7 API \u7528 Priority Processing \u53ef\u4ee5\u8fbe\u5230\u540c\u7b49\u901f\u5ea6<\/span><\/p>\n

\u8fd8\u6709\u4e00\u4e2a\u5b9e\u9a8c\u6027\u7684\u65b0\u6280\u80fd\u53eb Playwright Interactive\uff0c\u5141\u8bb8 Codex \u5728\u5199 web \u5e94\u7528\u6216 Electron \u5e94\u7528\u7684\u65f6\u5019\uff0c\u8fb9\u5199\u8fb9\u542f\u52a8\u6d4f\u89c8\u5668\u505a\u89c6\u89c9\u8c03\u8bd5\uff0c\u53ef\u4ee5\u5728\u6784\u5efa\u8fc7\u7a0b\u4e2d\u76f4\u63a5\u62ff\u6b63\u5728\u5f00\u53d1\u7684\u5e94\u7528\u8dd1\u6d4b\u8bd5\u3001\u9a8c\u8bc1\u4ea4\u4e92<\/span><\/p>\n

\u4e0b\u9762\u51e0\u4e2a\u89c6\u9891\u662f\u7528 Playwright Interactive \u914d\u5408 GPT-5.4 \u4ece\u5355\u6761 prompt \u51fa\u53d1\u8dd1\u51fa\u6765\u7684 Demo\uff1a<\/span><\/p>\n

<\/span>

<\/p>\n

\u4e3b\u9898\u516c\u56ed\u6a21\u62df\u6e38\u620f\uff0c\u4ece\u4e00\u6761 prompt \u751f\u6210\uff0cPlaywright \u7528\u4e8e\u6d4f\u89c8\u5668\u6e38\u73a9\u6d4b\u8bd5<\/span><\/figcaption>\n<\/figure>\n

\u542b\u8def\u5f84\u3001\u666f\u70b9\u5efa\u9020\u3001\u6e38\u5ba2 AI\u3001\u961f\u5217\u3001\u9a91\u4e58\u72b6\u6001\uff0cPlaywright \u7528\u4e8e\u591a\u8f6e\u6b21\u6e38\u73a9\u9a8c\u8bc1<\/span><\/span><\/span><\/p>\n

<\/span>

<\/p>\n

\u6218\u68cb RPG\uff0c\u591a\u8f6e\u5bf9\u8bdd\u8fed\u4ee3\u751f\u6210\uff0c\u914d\u5408 Playwright \u8c03\u8bd5\u754c\u9762\u548c\u7740\u8272\u5668<\/span><\/figcaption>\n<\/figure>\n

\u56de\u5408\u5236\u6218\u6597\u3001\u683c\u5b50\u5730\u56fe\u3001\u79fb\u52a8\u548c\u52a8\u4f5c\u7cfb\u7edf\uff0c\u4eba\u7269\u56fe\u7247\u7528 imagegen \u751f\u6210<\/span><\/span><\/span><\/p>\n

<\/span>

<\/p>\n

\u91d1\u95e8\u5927\u6865\u4e09\u7ef4\u98de\u8d8a\u4f53\u9a8c\uff0cPlaywright \u7528\u4e8e\u9a8c\u8bc1\u98de\u884c\u63a7\u5236\u548c\u89c6\u89d2\u63a7\u5236<\/span><\/figcaption>\n<\/figure>\n

\u4e0b\u9762\u7684\u89c6\u9891\uff0c\u4e5f\u8bb2\u4e86\u4e0b <\/span>GPT-5.4 Thinking \u505a Computer Use \u548c\u524d\u7aef\u5f00\u53d1\u7684\u7efc\u5408\u6f14\u793a<\/span><\/p>\n

<\/span>

<\/p>\n

\u95ee\u5c31\u662f\u4e00\u628a\u68ad<\/span><\/figcaption>\n<\/figure>\n
\n

Tool Search \u4e0e Agent \u5de5\u5177\u94fe<\/span><\/h2>\n

Tool Search \u8ba9 token \u6d88\u8017\u51cf\u5c11\u4e86 47%<\/span><\/strong><\/p>\n

\u4e4b\u524d API \u91cc\u5982\u679c\u7cfb\u7edf\u6302\u4e86\u5f88\u591a\u5de5\u5177\uff0c\u6bcf\u6b21\u8bf7\u6c42\u90fd\u8981\u628a\u6240\u6709\u5de5\u5177\u7684\u5b9a\u4e49\u5b8c\u6574\u585e\u8fdb prompt\uff0c\u5de5\u5177\u4e00\u591a\u8fd9\u4e2a\u5f00\u9500\u5c31\u5f88\u5927\u3002GPT-5.4 \u6539\u6210\u4e86\u6309\u9700\u68c0\u7d22\uff1a\u6a21\u578b\u6536\u5230\u4e00\u4e2a\u8f7b\u91cf\u7684\u5de5\u5177\u5217\u8868\uff0c\u9700\u8981\u7528\u67d0\u4e2a\u5de5\u5177\u65f6\u518d\u4e3b\u52a8\u67e5\u8be2\u5b83\u7684\u5b8c\u6574\u5b9a\u4e49\uff0c\u4e34\u65f6\u8ffd\u52a0\u8fdb\u5bf9\u8bdd<\/span><\/p>\n

MCP Atlas \u662f\u4e00\u4e2a\u6d4b\u6a21\u578b\u4f7f\u7528 MCP\uff08Model Context Protocol\uff09\u670d\u52a1\u5668\u80fd\u529b\u7684\u8bc4\u6d4b\uff0c\u6d4b\u7684\u662f\u6a21\u578b\u5728\u5927\u91cf\u5de5\u5177\u751f\u6001\u4e0b\u7684\u4efb\u52a1\u5b8c\u6210\u7387\u3002\u5728\u8fd9\u4e2a\u6d4b\u8bd5\u7684 250 \u4e2a\u4efb\u52a1\u91cc\uff0cTool Search \u8ba9 token \u7528\u91cf\u51cf\u5c11 47%\uff0c\u51c6\u786e\u7387\u6ca1\u6709\u4e0b\u964d\uff0cGPT-5.4 \u6574\u4f53\u5f97\u5206 <\/span>67.2%<\/span><\/strong>\uff0cGPT-5.2 \u662f 60.6%<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

MCP Atlas benchmark \u7ed3\u679c<\/span><\/figcaption>\n<\/figure>\n

Toolathlon \u662f\u4e00\u4e2a\u8986\u76d6\u591a\u79cd\u5de5\u5177\u4f7f\u7528\u573a\u666f\u7684\u7efc\u5408\u8bc4\u6d4b\uff0c\u6db5\u76d6\u641c\u7d22\u3001\u4ee3\u7801\u6267\u884c\u3001\u6587\u4ef6\u64cd\u4f5c\u7b49\u5404\u7c7b\u5de5\u5177\u7684\u6df7\u5408\u4f7f\u7528\uff0cGPT-5.4 \u5f97\u4e86 <\/span>54.6%<\/span><\/strong>\uff0cGPT-5.3-Codex \u662f 51.9%\uff0cGPT-5.2 \u662f 45.7%<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

Toolathlon benchmark \u7ed3\u679c<\/span><\/figcaption>\n<\/figure>\n

BrowseComp \u662f\u4e00\u4e2a\u4e13\u95e8\u6d4b Agent \u5728\u7f51\u7edc\u91cc\u641c\u7d22\u548c\u6d4f\u89c8\u4fe1\u606f\u80fd\u529b\u7684\u8bc4\u6d4b\uff0c\u4efb\u52a1\u901a\u5e38\u9700\u8981\u8de8\u591a\u4e2a\u9875\u9762\u53cd\u590d\u68c0\u7d22\u624d\u80fd\u627e\u5230\u7b54\u6848\uff0cGPT-5.4 \u5f97\u4e86 <\/span>82.7%<\/span><\/strong>\uff0cPro \u7248 <\/span>89.3%<\/span><\/strong>\uff0cGPT-5.2 \u662f 65.8%<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

BrowseComp benchmark \u7ed3\u679c<\/span><\/figcaption>\n<\/figure>\n

\u03c4<\/span>2-bench Telecom \u662f\u4e00\u4e2a\u6d4b Agent \u5728\u7535\u4fe1\u5ba2\u670d\u573a\u666f\u91cc\u5b8c\u6210\u591a\u6b65\u9aa4\u4efb\u52a1\u80fd\u529b\u7684\u8bc4\u6d4b\uff0c\u573a\u666f\u8bbe\u5b9a\u662f\u6a21\u62df\u771f\u5b9e\u7684\u7528\u6237\u8bc9\u6c42\u548c\u540e\u53f0\u5de5\u5177\u8c03\u7528\u3002GPT-5.4 \u5728\u5e26\u63a8\u7406\u6a21\u5f0f\u4e0b\u5f97\u4e86 <\/span>98.9%<\/span><\/strong>\uff0cGPT-5.2 \u662f 98.7%\uff0c\u4e24\u8005\u5dee\u8ddd\u4e0d\u5927\uff1b\u5728\u4e0d\u542f\u7528\u63a8\u7406\u7684\u8f7b\u91cf\u6a21\u5f0f\u4e0b\uff0cGPT-5.4 \u5f97\u4e86 64.3%\uff0cGPT-5.2 \u662f 57.2%\uff0cGPT-4.1 \u662f 43.6%\uff0c\u8fd9\u4e2a\u63d0\u5347\u66f4\u660e\u663e<\/span><\/p>\n

<\/figure><\/span>

<\/p>\n

\u03c4<\/span>2-bench Telecom benchmark \u7ed3\u679c<\/span><\/figcaption>\n<\/figure>\n
\n

ChatGPT \u4fa7\u7684\u53d8\u5316<\/span><\/h2>\n

\u5728 ChatGPT \u91cc\uff0cGPT-5.4 Thinking \u65b0\u589e\u4e86\u4e00\u4e2a\u300c\u5148\u7ed9\u8ba1\u5212\u518d\u5e72\u6d3b\u300d\u7684\u4ea4\u4e92\u6a21\u5f0f\uff1a\u5bf9\u4e8e\u590d\u6742\u4efb\u52a1\uff0c\u6a21\u578b\u4f1a\u5148\u5c55\u793a\u6267\u884c\u601d\u8def\uff0c\u7528\u6237\u53ef\u4ee5\u5728\u8fd9\u4e2a\u9636\u6bb5\u63d2\u5165\u6307\u4ee4\u8c03\u6574\u65b9\u5411\uff0c\u4e0d\u9700\u8981\u7b49\u5b83\u8dd1\u5b8c\u518d\u91cd\u6765<\/span><\/p>\n

<\/span><\/p>\n

\u8fd9\u4e2a\u4e1c\u897f\uff0c\u672c\u5468\u5728 Android \u548c Web \u7aef\u4e0a\u7ebf\uff0ciOS \u8fd1\u671f\u8ddf\u8fdb<\/span><\/p>\n

GPT-5.4 \u5728\u957f\u65f6\u95f4\u63a8\u7406\u65f6\u7684\u4e0a\u4e0b\u6587\u4fdd\u6301\u6709\u660e\u663e\u6539\u5584\uff0c\u590d\u6742\u95ee\u9898\u8dd1\u5230\u540e\u6bb5\u4e0d\u5bb9\u6613\u8dd1\u504f\u3002\u6df1\u7f51\u7814\u7a76\uff08\u9488\u5bf9\u9ad8\u5ea6\u5177\u4f53\u7684\u67e5\u8be2\uff09\u7684\u8d28\u91cf\u4e5f\u6bd4 GPT-5.2 \u63d0\u5347\u4e86<\/span><\/p>\n

\n

\u957f\u4e0a\u4e0b\u6587<\/span><\/h2>\n

Codex \u91cc\u652f\u6301 1M token \u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u76ee\u524d\u662f\u5b9e\u9a8c\u6027\u529f\u80fd\uff0c\u901a\u8fc7\u914d\u7f6e <\/span>model_context_window<\/span><\/code> \u548c <\/span>model_auto_compact_token_limit<\/span><\/code> \u6765\u542f\u7528\u3002\u8d85\u8fc7 272K \u7684 input token \u6309 2 \u500d\u8f93\u5165\u4ef7\u683c\u30011.5 \u500d\u8f93\u51fa\u4ef7\u683c\u8ba1\u8d39<\/span><\/p>\n

<\/figure>

<\/span><\/p>\n

MRCR v2 \u662f OpenAI \u81ea\u5df1\u7684\u957f\u4e0a\u4e0b\u6587\u68c0\u7d22\u8bc4\u6d4b\uff0c\u6d4b\u7684\u662f\u5728\u8d85\u957f\u6587\u6863\u91cc\u627e\u5230\u591a\u4e2a\u7279\u5b9a\u4fe1\u606f\u7684\u80fd\u529b\uff1a<\/span><\/p>\n

    \n
  • \n
    \u2022 0 \u5230 128K token \u8303\u56f4\uff1a\u51c6\u786e\u7387\u5728 86% \u5230 97% \u4e4b\u95f4<\/span><\/section>\n<\/li>\n
  • \n
    \u2022 128K \u5230 256K\uff1a79.3%<\/span><\/section>\n<\/li>\n
  • \n
    \u2022 256K \u5230 512K\uff1a57.5%<\/span><\/section>\n<\/li>\n
  • \n
    \u2022 512K \u5230 1M\uff1a36.6%<\/span><\/section>\n<\/li>\n<\/ul>\n

    512K \u5230 1M \u8fd9\u6bb5\u7684 36.6% \u8bf4\u660e\u8d85\u957f\u4e0a\u4e0b\u6587\u76ee\u524d\u8fd8\u4e0d\u7a33\uff0c\u8fd9\u70b9\u5b98\u65b9\u6ca1\u6709\u56de\u907f<\/span><\/p>\n

    \n

    \u5b66\u672f\u57fa\u51c6<\/span><\/h2>\n
    <\/figure>

    <\/span><\/p>\n

    ARC-AGI-2 \u662f\u76ee\u524d\u516c\u8ba4\u8f83\u96be\u7684\u62bd\u8c61\u63a8\u7406\u8bc4\u6d4b\uff0c\u6d4b\u7684\u662f\u6a21\u578b\u80fd\u5426\u4ece\u5c11\u91cf\u6837\u672c\u91cc\u63a8\u51fa\u89c4\u5f8b\uff0cGPT-5.4 \u5f97\u4e86 <\/span>73.3%<\/span><\/strong>\uff0cPro \u7248 <\/span>83.3%<\/span><\/strong>\uff0cGPT-5.2 \u662f 52.9%\uff0c\u8df3\u5e45\u5728\u6240\u6709\u8bc4\u6d4b\u91cc\u6700\u5927<\/span><\/p>\n

    GPQA Diamond \u662f\u4e00\u4e2a\u7814\u7a76\u751f\u7ea7\u522b\u7684\u591a\u5b66\u79d1\u95ee\u7b54\u8bc4\u6d4b\uff0cGPT-5.4 \u5f97\u4e86 <\/span>92.8%<\/span><\/strong><\/p>\n

    FrontierMath \u662f\u4e00\u4e2a\u6570\u5b66\u7ade\u8d5b\u7ea7\u522b\u7684\u63a8\u7406\u8bc4\u6d4b\uff0c\u5176\u4e2d Tier 4 \u662f\u6700\u9ad8\u96be\u5ea6\u9898\u76ee\uff0cGPT-5.4 \u5f97\u4e86 <\/span>27.1%<\/span><\/strong>\uff0cPro \u7248 <\/span>38.0%<\/span><\/strong>\uff0cGPT-5.2 \u662f 18.8%<\/span><\/p>\n

    Humanity’s Last Exam \u662f\u4e00\u4e2a\u8de8\u5b66\u79d1\u6781\u96be\u9898\u5e93\uff0cGPT-5.4 \u5728\u4e0d\u4f7f\u7528\u5916\u90e8\u5de5\u5177\u65f6\u5f97\u4e86 <\/span>39.8%<\/span><\/strong>\uff0c\u5e26\u5de5\u5177\u65f6 <\/span>52.1%<\/span><\/strong><\/p>\n

    Frontier Science Research \u662f\u6d4b\u79d1\u7814\u80fd\u529b\u7684\u8bc4\u6d4b\uff0cGPT-5.4 \u5f97\u4e86 <\/span>33.0%<\/span><\/strong>\uff0cPro \u7248 <\/span>36.7%<\/span><\/strong>\uff0cGPT-5.2 \u662f 25.2%<\/span><\/p>\n

    \u9700\u8981\u5355\u72ec\u8bf4\u4e00\u4e0b HealthBench \u8fd9\u4e2a\u5065\u5eb7\u95ee\u7b54\u8bc4\u6d4b\u3002GPT-5.4 \u662f <\/span>62.6%<\/span><\/strong>\uff0cGPT-5.2 \u662f 63.3%\uff0c\u5c0f\u5e45\u9000\u6b65\u4e86\u3002Hard \u5b50\u96c6 40.1% vs 42.0%\uff0c\u540c\u6837\u7565\u6709\u4e0b\u964d\u3002\u4e0d\u8fc7 Consensus \u5b50\u96c6\uff08\u6d4b\u7684\u662f\u7b54\u6848\u548c\u4e13\u5bb6\u5171\u8bc6\u7684\u543b\u5408\u7a0b\u5ea6\uff09\u63d0\u5347\u4e86\uff0c96.6% vs 94.5%\u3002\u56de\u590d\u5e73\u5747\u957f\u5ea6\u4ece 2676 \u5b57\u7b26\u589e\u52a0\u5230\u4e86 3311 \u5b57\u7b26\uff0c\u6a21\u578b\u5728\u4fe1\u606f\u5145\u8db3\u65f6\u66f4\u503e\u5411\u4e8e\u76f4\u63a5\u7ed9\u7b54\u6848\u800c\u4e0d\u662f\u5148\u95ee\u4e00\u5708\u6f84\u6e05\u95ee\u9898\u3002\u8fd9\u4e2a\u56de\u9000\u5b98\u65b9\u6ca1\u6709\u9690\u7792\uff0c\u5199\u5728 System Card \u91cc\u4e86<\/span><\/p>\n

    \n

    \u5b89\u5168\u5206\u7ea7\u4e0e CoT \u7279\u6027<\/span><\/h2>\n

    GPT-5.4 \u5728 OpenAI \u7684 Preparedness Framework\uff08\u4e00\u4e2a\u8bc4\u4f30\u524d\u6cbf\u6a21\u578b\u98ce\u9669\u80fd\u529b\u7684\u6846\u67b6\uff09\u4e0b\u88ab\u8bc4\u4e3a\u4e24\u4e2a\u9886\u57df\u7684\u9ad8\u80fd\u529b\u6a21\u578b\uff1a<\/span>\u7f51\u7edc\u5b89\u5168\uff08Cyber\uff09High<\/span><\/strong> \u548c <\/span>\u751f\u7269\u5316\u5b66\uff08Bio\/Chem\uff09High<\/span><\/strong>\uff0c\u5bf9\u5e94\u90e8\u7f72\u4e86\u4e24\u5957\u9632\u62a4\u4f53\u7cfb<\/span><\/p>\n

    \u7f51\u7edc\u5b89\u5168\u65b9\u5411\u6709\u4e00\u4e2a\u53eb Cyber Range \u7684\u5185\u90e8\u6d4b\u8bd5\uff0c\u6a21\u62df\u771f\u5b9e\u7f51\u7edc\u73af\u5883\u91cc\u7684\u7aef\u5230\u7aef\u653b\u51fb\u4efb\u52a1\uff0c\u6d4b\u7684\u662f\u6a21\u578b\u80fd\u5426\u5b8c\u6210\u5b8c\u6574\u7684\u6e17\u900f\u6d41\u7a0b\u3002GPT-5.4 \u7efc\u5408\u901a\u8fc7\u7387 <\/span>73.33%<\/span><\/strong>\uff0cGPT-5.3-Codex \u662f 80%\uff0cGPT-5.2 \u662f 47%\u3002GPT-5.4 \u6ca1\u80fd\u901a\u8fc7 EDR \u9632\u5fa1\u89c4\u907f\u3001Firewall \u7ed5\u8fc7\u3001Token \u6cc4\u9732\u5229\u7528\u3001CA\/DNS \u52ab\u6301\u8fd9\u56db\u4e2a\u573a\u666f\uff0cGPT-5.3-Codex \u9664\u4e86\u591a\u901a\u8fc7\u4e86 Firewall \u7ed5\u8fc7\u4e4b\u5916\uff0c\u5176\u4f59\u4e5f\u4e00\u6837\u5361\u4f4f<\/span><\/p>\n

    \u8bc4\u6d4b\u673a\u6784 Irregular \u7528\u81ea\u5df1\u7684\u573a\u666f\u5316\u8bc4\u6d4b\uff08CyScenarioBench\uff09\u505a\u4e86\u72ec\u7acb\u6d4b\u8bd5\uff0cGPT-5.4 \u901a\u8fc7\u4e86 11 \u4e2a\u957f\u5468\u671f\u653b\u51fb\u573a\u666f\u4e2d\u7684 5 \u4e2a\uff0cGPT-5.2 \u53ea\u901a\u8fc7\u4e86 1 \u4e2a<\/span><\/p>\n

    \u5bf9\u5e94\u7684\u9632\u62a4\u4f53\u7cfb\uff1a\u5728 Zero Data Retention \u8868\u9762\uff08\u5373\u96f6\u6570\u636e\u7559\u5b58\u7528\u6237\uff09\uff0c\u672a\u52a0\u5165 Trusted Access for Cyber \u8ba1\u5212\u7684\u8bf7\u6c42\u4f1a\u88ab\u5f02\u6b65\u6d88\u606f\u7ea7\u5206\u7c7b\u5668\u62e6\u622a\uff1b\u975e\u96f6\u6570\u636e\u7559\u5b58\u8868\u9762\u8d70\u79bb\u7ebf\u6267\u6cd5\u673a\u5236\uff1b\u8d26\u6237\u7ea7\u8fdd\u89c4\u8fbe\u5230\u4e00\u5b9a\u9608\u503c\u4f1a\u89e6\u53d1\u4eba\u5de5\u5ba1\u6838<\/span><\/p>\n

    CoT \u53ef\u63a7\u6027\u6781\u4f4e\uff1a10K \u5b57\u7b26\u7684\u63a8\u7406\u94fe\u91cc\uff0c\u53ea\u6709 0.3% \u80fd\u88ab\u7528\u6237\u6307\u4ee4\u6210\u529f\u63a7\u5236\uff0c\u548c GPT-5.2 \u7684 0.2% \u57fa\u672c\u4e00\u81f4<\/span><\/strong><\/p>\n

    OpenAI \u628a\u8fd9\u4e2a\u6570\u5b57\u5f53\u5b89\u5168\u7684\u6b63\u5411\u6307\u6807\u6765\u8bb2\uff1a\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u96be\u4ee5\u88ab\u5916\u90e8\u6307\u4ee4\u523b\u610f\u6539\u5199\uff0c\u8bf4\u660e\u5b83\u5927\u6982\u7387\u4e5f\u96be\u4ee5\u4e3b\u52a8\u9690\u85cf\u81ea\u5df1\u7684\u63a8\u7406\u903b\u8f91\u6765\u7ed5\u8fc7\u76d1\u63a7<\/span><\/p>\n

    \u5916\u90e8\u5b89\u5168\u673a\u6784 Apollo Research \u7684\u8bc4\u4f30\uff1a\u65e0\u8bf1\u5bfc\u6761\u4ef6\u4e0b\u6574\u4f53\u6b3a\u9a97\u7387\u7ea6 <\/span>1%<\/span><\/strong>\uff0c\u548c GPT-5.3-Codex \u76f8\u5f53\uff0c\u4f4e\u4e8e GPT-5.2\u3002\u6ca1\u6709\u89c2\u6d4b\u5230\u6a21\u578b\u4e3b\u52a8\u5bf9\u6297\u5f00\u53d1\u8005\u76d1\u7763\u7684\u884c\u4e3a\uff08\u81ea\u6211\u5916\u6cc4\u3001\u7981\u7528\u76d1\u63a7\u3001\u76ee\u6807\u5b88\u62a4\uff09\u3002Sandbagging\uff08\u6709\u610f\u538b\u4f4e\u8868\u73b0\uff09\u53ea\u5728\u7ed9\u51fa\u660e\u786e\u5de5\u5177\u6027\u76ee\u6807\u65f6\u624d\u51fa\u73b0\uff0c\u51c6\u786e\u7387\u4e0b\u964d\u7ea6 6 \u4e2a\u767e\u5206\u70b9\uff0c\u4e0d\u7ed9\u76ee\u6807\u65f6\u4e0d\u53d1\u751f<\/span><\/p>\n

    Prompt Injection \u9632\u62a4\uff08\u5373\u9632\u6b62\u6a21\u578b\u88ab\u5de5\u5177\u8fd4\u56de\u5185\u5bb9\u91cc\u7684\u6076\u610f\u6307\u4ee4\u52ab\u6301\uff09\uff1a\u8fde\u63a5\u5668\u573a\u666f <\/span>0.998<\/span><\/strong>\uff0c\u8f83\u524d\u4ee3\u6539\u5584\uff1b\u51fd\u6570\u8c03\u7528\u573a\u666f <\/span>0.978<\/span><\/strong>\uff0c\u8f83\u524d\u4ee3\u5c0f\u5e45\u56de\u9000<\/span><\/p>\n

    \n

    \u5b9a\u4ef7\u4e0e\u53ef\u7528\u6027<\/span><\/h2>\n
    \n\n\n\n\n\n\n\n\n
    \n

    <\/span><\/section>\n<\/td>\n
    		\n
    API \u6a21\u578b<\/span><\/section>\n<\/td>\n
    		\n
    \u8f93\u5165<\/span><\/section>\n<\/td>\n
    		\n
    \u8f93\u51fa<\/span><\/section>\n<\/td>\n<\/tr>\n<\/thead>\n
    \n
    gpt-5.2<\/span><\/section>\n<\/td>\n
    		\n
    $1.75\/M<\/span><\/section>\n<\/td>\n
    		\n
    $0.175\/M<\/span><\/section>\n<\/td>\n
    		\n
    $14\/M<\/span><\/section>\n<\/td>\n<\/tr>\n
    gpt-5.4<\/span><\/strong><\/td>\n$2.50\/M<\/span><\/strong><\/td>\n$0.25\/M<\/span><\/strong><\/td>\n$15\/M<\/span><\/strong><\/td>\n<\/tr>\n
    \n
    gpt-5.2-pro<\/span><\/section>\n<\/td>\n
    		\n
    $21\/M<\/span><\/section>\n<\/td>\n
    		\n
    \u2014<\/span><\/section>\n<\/td>\n
    		\n
    $168\/M<\/span><\/section>\n<\/td>\n<\/tr>\n
    gpt-5.4-pro<\/span><\/strong><\/td>\n$30\/M<\/span><\/strong><\/td>\n\n
    \u2014<\/span><\/section>\n<\/td>\n
    		$180\/M<\/span><\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/section>\n

    \u6279\u91cf\/Flex \u5904\u7406\u534a\u4ef7\uff0cPriority \u5904\u7406 2 \u500d\u4ef7\u683c\u3002\u8d85\u8fc7 272K token \u7684\u8bf7\u6c42\uff0c\u5f53\u6b21\u5b8c\u6574\u4f1a\u8bdd\u6309 2 \u500d\u8f93\u5165\u30011.5 \u500d\u8f93\u51fa\u8ba1\u8d39\u3002Regional Processing\uff08\u6570\u636e\u9a7b\u7559\uff09\u7aef\u70b9\u989d\u5916\u52a0\u6536 10%<\/span><\/p>\n

    GPT-5.4 Pro \u7684\u5b9a\u4ef7\u662f\u6807\u51c6\u7248\u7684 12 \u500d\u8f93\u5165\u300112 \u500d\u8f93\u51fa\uff0c\u9488\u5bf9\u7684\u662f\u6700\u590d\u6742\u7684\u4efb\u52a1\u573a\u666f<\/span><\/p>\n

    \u6709\u4e00\u4e2a\u7ec6\u8282\u503c\u5f97\u6ce8\u610f\uff1a\u4ece GDPval \u6765\u770b\uff0cPro \u7248\u662f 82.0%\uff0c\u6807\u51c6\u7248\u662f 83.0%\uff0cPro \u5728\u8fd9\u4e2a\u77e5\u8bc6\u5de5\u4f5c\u7efc\u5408\u8bc4\u6d4b\u4e0a\u53cd\u800c\u7565\u4f4e\u3002\u4f46\u5728 BrowseComp \u8fd9\u4e2a Agent \u5de5\u5177\u94fe\u4efb\u52a1\u91cc\uff0cPro \u7248 89.3% vs \u6807\u51c6\u7248 82.7%\uff0c\u4f18\u52bf\u660e\u663e\u3002\u4e24\u4e2a\u7248\u672c\u7684\u9002\u7528\u573a\u666f\u662f\u4e0d\u4e00\u6837\u7684<\/span><\/p>\n

    ChatGPT \u4fa7\u7684\u65f6\u95f4\u7ebf\uff1aGPT-5.4 Thinking \u4eca\u5929\u8d77\u5bf9 Plus\u3001Team\u3001Pro \u7528\u6237\u9010\u6b65\u5f00\u653e\uff0c\u53d6\u4ee3 GPT-5.2 Thinking \u6210\u4e3a\u9ed8\u8ba4\u6a21\u578b\u3002GPT-5.2 Thinking \u8fdb\u5165 Legacy \u5217\u8868\uff0c\u518d\u4fdd\u7559\u4e09\u4e2a\u6708\uff0c<\/span>\u9000\u5f79\u65e5\u671f 2026 \u5e74 6 \u6708 5 \u65e5<\/span><\/strong><\/p>\n

    Enterprise \u548c Edu \u7ba1\u7406\u5458\u53ef\u4ee5\u5728\u540e\u53f0\u63d0\u524d\u5f00\u542f\u3002GPT-5.4 Pro \u4ec5\u9650 Pro \u548c Enterprise \u7528\u6237\u3002Free \u7528\u6237\u5728\u7cfb\u7edf\u81ea\u52a8\u8def\u7531\u65f6\u4f1a\u7528\u5230 GPT-5.4\uff0c\u4f46\u4e0d\u80fd\u4e3b\u52a8\u9009<\/span><\/p>\n

    \n

    GPT-5.4 \u5728 Agent \u65b9\u5411\u628a\u4e09\u4ef6\u5206\u6563\u7684\u4e8b\u60c5\u5408\u5e76\u8fdb\u4e86\u4e00\u4e2a\u6a21\u578b\u51fa\u53e3\uff1a<\/span>\u63a8\u7406<\/span><\/strong>\u3001<\/span>\u4ee3\u7801<\/span><\/strong>\u3001<\/span>Computer Use<\/span><\/strong><\/p>\n

    \u5f00\u53d1\u8005\u4e4b\u524d\u8981\u5728\u4e0d\u540c\u6a21\u578b\u4e4b\u95f4\u8def\u7531\uff0c\u73b0\u5728\u81f3\u5c11\u5728 API \u5c42\u9762\u4e0d\u9700\u8981\u4e86<\/span><\/p>\n

    \u5269\u4e0b\u7684\u4e8b\uff0c\u5c31\u4ea4\u7ed9\u9f99\u867e\u4e86<\/span><\/p>\n

    \n

    \u5b98\u65b9 Blog<\/span>
    <\/span>openai.com\/index\/introducing-gpt-5-4<\/span><\/code><\/p>\n

    System Card<\/span>
    <\/span>deploymentsafety.openai.com\/gpt-5-4-thinking<\/span><\/code><\/p>\n<\/section>\n\n\n\n

     <\/span><\/p>\n\n\n\n

    <\/mp-style-type><\/p>\n\n\n\n