豆包再扔王炸！2.0发布：推理成本降一个数量级，正面对标GPT-5和Gemini 3

Core Viewpoint - ByteDance's Doubao model has officially entered the 2.0 phase, offering a systematic upgrade that maintains performance comparable to GPT-5.2 and Gemini 3 Pro while reducing reasoning costs by approximately an order of magnitude, providing a competitive solution for complex tasks in large-scale production environments [2][12]. Group 1: Model Features and Performance - The Doubao 2.0 series includes Pro, Lite, Mini general-purpose agent models, and a specialized Code model, with the flagship Doubao 2.0 Pro achieving top scores in visual understanding benchmarks and winning gold medals in math Olympiads (IMO, CMO) and programming competitions (ICPC) [2][9]. - Doubao 2.0 has significantly upgraded its multimodal capabilities, excelling in tasks such as visual reasoning, perception, spatial reasoning, and long-context understanding [2]. - In dynamic scene understanding, Doubao 2.0 leads in key assessments like TVBench and surpasses human scores in EgoTempo, demonstrating stable capture of changes, actions, and rhythms [4]. - In long video scenarios, Doubao 2.0 outperforms other top models in most evaluations and excels in real-time Q&A video benchmark tests [5]. Group 2: Cost Efficiency and Application - Doubao 2.0 Pro has enhanced long-tail domain knowledge, scoring higher than GPT-5.2 on SuperGPQA and ranking first on HealthBench, with overall performance comparable to Gemini 3 Pro and GPT-5.2 in scientific fields [8]. - The model achieved a top score of 54.2 on HLE-text (Human Last Exam) and demonstrated excellent performance in tool invocation and instruction-following tests [10]. - The significant cost advantage of Doubao 2.0, with token pricing reduced by about an order of magnitude, will be crucial in large-scale reasoning and long-chain generation scenarios [12]. Group 3: Development and Integration - ByteDance has built an intelligent customer service agent on Feishu based on the OpenClaw framework and Doubao 2.0 Pro model, capable of handling customer dialogues and proactively seeking human assistance when faced with challenges [13][14]. - The Doubao 2.0 Code model is optimized for programming scenarios, enhancing code library interpretation and application generation capabilities, and has been integrated into the TRAE product [15][16]. - Developers using TRAE with Doubao 2.0 Code can create interactive projects with minimal prompts, showcasing the model's efficiency in project development [16][17]. - Doubao 2.0 Pro is now available to end-users on the Doubao App, desktop, and web versions, while API services for enterprises and developers have been launched on the Volcano Engine [18].