Grok 4.5 Released
Thursday, July 9, 2026 · Time not confirmed

Details
- Brand / Manufacturer
- SpaceXAI
- Product type
- AI model
- Model
- Grok 4.5
- Key specifications
- Served at 80 TPS; $2 per million input tokens, $6 per million output tokens; trained across tens of thousands of NVIDIA GB300 GPUs
- Region Availability
- Global
Release information
The model rolled out simultaneously across Grok.com, X, iOS, Android, the SpaceXSI API console, and Cursor on all plans. Free access to Grok 4.5 is offered for a limited time within Grok Build and Cursor.
Grok 4.5 is an artificial intelligence model developed by SpaceXSI, designed specifically for software engineering, scientific inquiry, math, and general knowledge tasks. Built to execute complex multi-step workflows, it operates as an automated agent capable of generating working applications, detailed financial models, and presentation diagrams from single user prompts.
The model serves as the primary engine within the Grok Build platform and integrates directly into desktop productivity applications. It handles end-to-end assignments such as data analysis, prose composition, and document summarization across both desktop and mobile environments.
What we know
Grok 4.5 was developed by SpaceXSI in close collaboration with the coding platform startup Cursor. The architecture focuses on agentic tasks, science, mathematics, and advanced software development. By optimizing per-token reasoning, the system reduces the total number of computational steps required to resolve complex technical problems.
The model targets complex end-to-end tasks without requiring step-by-step guidance from human users. It is capable of generating fully functional code bases, interactive web applications, and detailed documentation from concise initial prompts. In addition to desktop web access, it functions across mobile platforms and standalone developer interfaces.
Engineering performance and benchmarks
SpaceXSI evaluated Grok 4.5 across several recognized software engineering benchmarks to measure real-world performance. On the SWE Bench Pro evaluation, the model resolved coding challenges using an average of 15,954 output tokens. This represents roughly a fourfold improvement in token efficiency compared to competing top-tier architectures.
Additional benchmark results highlights include a 62.0 percent pass rate on DeepSWE 1.0 and a 29.0 percent resolution rate on SWE Marathon. It achieved an 83.3 percent score on Terminal Bench 2.1, showcasing strong terminal execution and command-line reasoning skills. These evaluations demonstrate high intelligence per output token across technical environments.
Office productivity and application tools
Beyond core programming, Grok 4.5 is integrated directly into Microsoft Office application plugins. Dedicated extensions allow the model to operate within Word, PowerPoint, Excel, and Outlook. Users can trigger document generation, presentation outline designs, and email drafting directly within their existing workspace tools.
When executing financial and spreadsheet tasks in Excel, the model generates multi-sheet formulas, incorporates web research, and inserts explanatory notes for future reference. Within PowerPoint, it uses native shape components to design custom diagrams and slide structures. It also serves as the underlying default engine for automated tasks inside Grok Build.
Training architecture and infrastructure
Training for Grok 4.5 relied on asynchronous reinforcement learning workflows spanning hundreds of thousands of specialized technical tasks. Automated and model-based grading systems maintained quality control throughout the optimization process. The infrastructure allowed agentic rollout processes to execute continuously for many hours without interrupting training cycles.
Data curation efforts focused on deduplication, quality scoring, and domain-specific selection to create a high-signal data mixture. The resulting model exhibits enhanced context retention, permitting long multi-turn conversations without losing track of previous instructions. This training methodology ensures reliable reasoning during extended logic problems and continuous project builds.
Integration options and subscription tiers
Developers can access Grok 4.5 via the SpaceXSI API console or third-party routing platforms such as OpenRouter. Integration is supported directly within Cursor across all subscriber tiers, alongside native terminal installation scripts. Developers can execute API requests for automated code debugging, bug identification, and instant code generation.
Individual consumers access the model through various tier levels including Plus, Premium, Super, and Ultra subscription plans. Higher subscription tiers provide up to fifty times the baseline usage allowance for chat, voice, image generation, and cloud agent automations. Enterprise setups also include administrative spend controls, audit logging, and team-wide security settings.
Frequently asked questions
- When does Grok 4.5 2026 release?
- Grok 4.5 2026 is scheduled to release on Thursday, July 9, 2026 (exact time not yet confirmed).
- Where is Grok 4.5 available for developers?
- Developers can access Grok 4.5 through the SpaceXSI API console, third-party hosts such as OpenRouter, and directly within the Cursor development platform across all subscription plans.
- What features does Grok 4.5 provide inside Microsoft Office software?
- Plugins for Word, PowerPoint, Excel, and Outlook allow Grok 4.5 to build multi-sheet financial models, design slide diagrams using native shapes, draft document prose, and summarize emails.
- What benchmark results did Grok 4.5 achieve on coding tests?
- Grok 4.5 scored 62.0% on DeepSWE 1.0, 83.3% on Terminal Bench 2.1, 64.7% on SWE Bench Pro, and 29.0% on SWE Marathon.
- How does Grok 4.5 handle prompt-based application development?
- Grok 4.5 can build complete working web applications, single-page landing sites, and playable 3D browser games in Three.js from a single text prompt.
Last updated: 11/14/2026





