Grok 4.5 模型卡
Grok 4.5 的官方模型卡，記錄了該模型的能力基準測試與安全評估。
Grok
Our most intelligent model for long-running agentic work across coding and knowledge work.
Announcements
推出 Grok 4.5
這是我們最聰明的模型，也是首度打造、適用於軟體工程以外領域的模型。
Cursor 與 SpaceX 合作推進模型訓練
Cursor 正與 SpaceX 合作，加速我們的模型訓練進程。
Built for coding and knowledge work alike
Long-horizon problemsLong-horizon
problems
Grok 4.5 runs hard problems for hours: investigating, using tools creatively, recovering from dead ends, and verifying results. Built for work that needs sustained reasoning and multiple tool calls before a conclusion is reached.
Ambitious engineeringAmbitious
engineering
Grok 4.5 is strong at large migrations, multi-service systems, and ultra-long-horizon engineering projects. It reads the codebase, edits across files, runs tests, and keeps going until the patch holds up.
In-depth knowledge workIn-depth
knowledge work
Beyond code, Grok 4.5 excels at deep research, analysis, and professional deliverables across data science, finance, legal, and other forms of knowledge work. Give it a multi-step project that requires research, analysis, and synthesis and review a polished, final deliverable in turn.
Research and investigationResearch and
investigation
Grok 4.5 runs multi-step research with tools: searching, synthesizing sources, and verifying its claims for absolute honesty and accuracy. It's designed for a measured, honest approach with the lowest single-turn hallucination rate on the work it performs.
Built for more than software engineering
Grok 4.5 handles difficult, long-running tasks that require creatively using tools to solve problems, whether in software engineering, data science, finance, legal work, or anything else you do on a computer.
It solves multistep tasks in under half the steps of comparable frontier models. Cursor subscription plans for individuals and teams include significant usage of the model, with double usage for the first week.
A strong foundation
Grok 4.5 is a mixture-of-experts model that we trained jointly with SpaceXAI. Training included trillions of tokens of Cursor data that capture real developer-agent interactions, so the model learns both from existing software and from how agents work inside a codebase.
The training mix is broader than Composer 2.5. It draws on high-quality STEM tasks, research papers, and other knowledge work, then adds reinforcement learning on difficult, realistic problems that teach the model to investigate, recover from mistakes, and verify results.
Benchmarks
| Grok 4.5 | Opus 4.8 | GPT-5.5 | Composer 2.5 | Fable 5 | |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 83.3% | 78.9% | 83.4% | 73.0% | 84.3% |
| SWE-Bench Multilingual | 78.0% | 84.4% | 77.8% | 71.6% | — |
| DeepSWE 1.0Artificial Analysis | 62.0% high | 55.8% max | 64.3% xhigh | 18.0% | 66.1% max |
| SWE-Bench Pro | 64.7% high | 69.2% max | 58.6% xhigh | 54.0% | 80.3% |
SWE-Bench Pro and Terminal-Bench show self-reported scores for third-party models. For SWE-Bench multilingual, the GPT-5.5 score comes from our internal run.
(Above) Grok 4.5 results across SWE-Bench Pro, Terminal-Bench, SWE-Bench multilingual, and CursorBench.
Available everywhere you work
Grok 4.5 is available today in Cursor across desktop, web, iOS, CLI, and our SDK.
Interactive demo with multiple windows showing Cursor's AI-powered features. The interface is displayed over a subtle, solid brand background.
Set up Datadog APM for all API routes
Explored 12 files, 3 searches
I'll add the Datadog SDK, instrument all API routes with tracing, and set up error tracking.
Worked for 7m 42s
Processed screen recording
Done — here's the Datadog dashboard showing all instrumented routes.
Acme Labs
Summary
Integrated Datadog APM across all 24 API routes with error tracking and custom metrics.
Add a follow up...
This element contains an interactive demo for sighted users. It's a demonstration of Cursor integrated within Slack, showing AI-powered assistance inside team communication. The interface is displayed over a subtle, solid brand background.
FAQ
Grok 4.5 is Cursor's most intelligent model, and the first built for more than software engineering. It's a mixture-of-experts model trained jointly with SpaceXAI that handles difficult, long-running tasks requiring creative tool use, across software engineering, data science, finance, legal work, and other computer-based tasks. It solves multistep tasks in under half the steps of comparable frontier models.
Grok 4.5 is available today in Cursor across desktop, web, iOS, CLI, and our SDK. You can also build your own agents on top of Grok 4.5 with our SDK docs.
Grok 4.5 is part of the first-party models pool on individual and team plans, alongside Auto and Composer 2.5, and included usage is doubled through July 21, 2026. The base model is priced at $2/M input and $6/M output tokens; the Fast variant is $4/M input and $18/M output tokens. Grok 4.5 is not yet available in the European Union in accordance with the EU AI Act. See the model docs for full details.