AI News AI资讯 1d ago Updated 1d ago 更新于 1天前 49

SpaceXAI Releases Grok 4.5, a Cursor-Trained Model for Coding, Agentic Tasks, and Knowledge Work at $2/M Input SpaceXAI发布Grok 4.5:一款针对编码、智能体任务和知识工作的Cursor训练模型,输入成本为2美元/百万Token

SpaceXAI released Grok 4.5, a general-purpose model optimized for coding, agentic tasks, and knowledge work, trained in collaboration with Cursor. The model demonstrates superior token efficiency, using approximately 4.2x fewer output tokens than Opus 4.8 (max) on SWE Bench Pro, resulting in lower latency and cost. Training leveraged tens of thousands of NVIDIA GB300 GPUs with a focus on high-quality data curation and asynchronous reinforcement learning for multi-step software engineering tasks. SpaceXAI发布Grok 4.5,定位为目前最智能的模型,专注于编码、代理任务和知识工作。 该模型与Cursor编辑器联合训练,在SWE Bench Pro上展现出极高的Token效率,输出Token数仅为Opus 4.8的约1/4.2。 硬件训练依托数万张NVIDIA GB300 GPU,采用大规模强化学习及异步训练技术,强调每Token的智能性。 定价为输入$2/M、输出$6/M,推理速度达80 TPS,并在Harvey法律代理基准测试中排名第一。 模型ID为grok-4.5,已上线Grok Build、Cursor全计划及SpaceXAI控制台,欧盟地区预计7月中旬开放。

75
Hot 热度
65
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • SpaceXAI released Grok 4.5, a general-purpose model optimized for coding, agentic tasks, and knowledge work, trained in collaboration with Cursor.
  • The model demonstrates superior token efficiency, using approximately 4.2x fewer output tokens than Opus 4.8 (max) on SWE Bench Pro, resulting in lower latency and cost.
  • Training leveraged tens of thousands of NVIDIA GB300 GPUs with a focus on high-quality data curation and asynchronous reinforcement learning for multi-step software engineering tasks.
  • Grok 4.5 achieves top-tier performance on benchmarks like Harvey’s Legal Agent Benchmark (#1) and Terminal Bench 2.1, while offering competitive pricing at $2/M input and $6/M output tokens.

Why It Matters

This release highlights a critical industry shift toward optimizing for token efficiency and cost-effectiveness in agentic workflows, rather than solely pursuing raw benchmark accuracy. For AI practitioners, the integration with Cursor and strong performance in legal and coding domains suggests that specialized, efficient models may offer better ROI for enterprise automation than larger, less efficient alternatives. The emphasis on asynchronous RL and per-token intelligence provides a blueprint for training models that are not just smart, but also economically viable for high-volume, multi-step tasks.

Technical Details

  • Training Infrastructure & Methodology: Utilized tens of thousands of NVIDIA GB300 GPUs with advanced stability techniques. The process involved rigorous data filtering (deduplication, quality scoring) and scaled reinforcement learning focused on "per-token intelligence" across hundreds of thousands of multi-step software engineering tasks.
  • Performance Metrics: Achieved an 83.3% pass rate on Terminal Bench 2.1 and a 64.7% resolve rate on SWE Bench Pro. Notably, it required only 15,954 average output tokens to resolve SWE Bench Pro tasks compared to 67,020 for Opus 4.8 (max).
  • Benchmark Leadership: Ranked #1 on Harvey’s Legal Agent Benchmark and showed competitive results against GPT 5.5 (xhigh) and Opus 4.8 (max) across DeepSWE and Terminal Bench harnesses.
  • Deployment & Accessibility: Available via the SpaceXAI console (API ID: grok-4.5), integrated into Cursor on all plans, and serves as the default model in Grok Build. It is currently unavailable in the EU, with expected availability in mid-July.

Industry Insight

  • Cost Optimization Strategy: The significant reduction in output tokens suggests that future model evaluations should prioritize efficiency metrics alongside accuracy, as lower token usage directly translates to reduced operational costs for agentic applications.
  • Ecosystem Integration: The deep integration with development tools like Cursor indicates that the most successful AI models will be those embedded directly into developer workflows, enabling seamless transitions between ideation, coding, and debugging.
  • Specialization vs. Generalization: While positioned as a general-purpose model, its strong showing in legal and coding benchmarks implies that targeted fine-tuning on specific high-value domains (like law or software engineering) yields competitive advantages over purely broad-spectrum models.

TL;DR

  • SpaceXAI发布Grok 4.5,定位为目前最智能的模型,专注于编码、代理任务和知识工作。
  • 该模型与Cursor编辑器联合训练,在SWE Bench Pro上展现出极高的Token效率,输出Token数仅为Opus 4.8的约1/4.2。
  • 硬件训练依托数万张NVIDIA GB300 GPU,采用大规模强化学习及异步训练技术,强调每Token的智能性。
  • 定价为输入$2/M、输出$6/M,推理速度达80 TPS,并在Harvey法律代理基准测试中排名第一。
  • 模型ID为grok-4.5,已上线Grok Build、Cursor全计划及SpaceXAI控制台,欧盟地区预计7月中旬开放。

为什么值得看

Grok 4.5通过“联合训练”模式展示了垂直领域工具(如Cursor)对大模型能力的显著增强,为开发者提供了优化编码和代理任务的新范式。其卓越的Token效率和成本控制(相比竞品大幅减少输出Token),使其在企业级应用和高并发场景下具备极高的商业吸引力。

技术解析

  • 训练架构与数据:模型在数万张NVIDIA GB300 GPU集群上训练,数据经过严格的去重、质量评分和领域筛选。强化学习阶段覆盖数十万项任务,重点在于多步软件工程,采用自动化与模型结合的评价方法,支持长时间异步训练。
  • 性能基准对比:在DeepSWE 1.0和Terminal Bench 2.1等编码基准中表现强劲,虽在部分榜单略低于Fable (max),但在SWE Bench Pro的解决率上达到64.7%,且显著优于Opus 4.8 (55.75%)和GPT 5.5 (58.6%)。
  • Token效率与成本:核心优势在于效率,解决相同任务平均仅需15,954个输出Token,而Opus 4.8需67,020个,效率提升约4.2倍。这直接降低了延迟和API调用成本。
  • 部署与服务:提供80 TPS的服务吞吐量,模型ID为grok-4.5。支持通过标准REST API调用,并集成于Grok Build CLI和Cursor编辑器中,便于开发者快速接入。

行业启示

  • 工具链融合趋势:通用大模型与专用开发工具(如IDE、编辑器)的深度耦合将成为提升特定领域(如编程、法律)任务表现的关键路径,未来模型训练将更依赖真实工作流数据。
  • 效率优先于单纯规模:在推理成本日益敏感的背景下,降低Token消耗和提升单次请求的信息密度(Per-Token Intelligence)比单纯增加参数规模更具商业价值,企业应关注模型的“性价比”而非仅看基准分数。
  • 异步与长期代理能力:支持长时间异步训练和复杂多步任务的RL策略,表明AI Agent正从简单的问答向需要持久化思考和执行复杂工程任务的方向演进,这对基础设施的稳定性和调度提出了新要求。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Code Generation 代码生成 Agent Agent Product Launch 产品发布 Training 训练