Grok 4.5 Release Notes: Advanced Coding and Agentic Capabilities

Grok 4.5 delivers high-performance reasoning for engineering and agentic tasks

SpaceXAI has released Grok 4.5, a model specifically engineered to excel in coding, agentic workflows, and knowledge work. The model is designed for real-world engineering excellence, combining intelligent reasoning with high efficiency to outperform comparable models in specific technical benchmarks.

Technical Benchmarks and Engineering Performance

Grok 4.5 demonstrates strong performance across several software engineering benchmarks, particularly in resolution rates and efficiency.

• SWE Marathon: Grok 4.5 leads with a 29.0% resolution rate (pass@1), surpassing Opus 4.8 (26.0%) and Fable (24.0%).
• Terminal Bench 2.1: The model achieves a score of 83.3%, placing it closely behind Fable (84.3%) and GPT 5.5 (83.4%).
• SWE Bench Pro: Grok 4.5 achieves a 64.7% resolve rate, outperforming GPT 5.5 (58.6%) and Opus 4.7 (64.3%).
• DeepSWE 1.0: It scores 62.0%, trailing Fable (66.1%) and GPT 5.5 (64.31%).
• DeepSWE 1.1: It scores 53%, following Fable (70%) and GPT 5.5 (67%).

Training Methodology and Infrastructure

Grok 4.5 was trained on tens of thousands of NVIDIA GB300 GPUs using a combination of large-scale stability techniques and rigorous data curation.

• Data Strategy: The training process emphasized deduplication, quality scoring, and domain-focused selection to ensure a high-signal, high-coverage data mixture across coding, science, engineering, and math.
• Reinforcement Learning (RL): SpaceXAI scaled RL with a focus on per-token intelligence, covering hundreds of thousands of tasks centered on multi-step software engineering. This was achieved through automated and model-based grading.
• Asynchronous Training: The training stack supports highly asynchronous rollouts, allowing agentic processes to run for several hours while learning continues across the GPU cluster.

Speed, Efficiency, and Cost

Grok 4.5 is optimized for low latency and cost-effectiveness, offering speeds of 80 tokens per second (TPS).

• Token Efficiency: On the SWE Bench Pro task, Grok 4.5 resolves tasks using an average of 15,954 output tokens, which is approximately 4.2x fewer tokens than Opus 4.8 (max) at 67,020 tokens.
• Pricing: The model is priced at $2 per million input tokens and $6 per million output tokens.

Integration and Productivity Tools

Beyond raw coding, Grok 4.5 is integrated into productivity suites and development environments:

• Grok Build: Now the default model, enabling the creation of complex Excel models with web research, multi-sheet formulas, and internal notes.
• Office Integration: The model can use native PowerPoint shapes to build diagrams and write prose in Word.
• Availability: Grok 4.5 is available via the SpaceXAI console, Cursor (on all plans), and Grok Build. It is currently unavailable in the EU, with availability expected in mid-July.

Coding Capabilities

Grok 4.5 is proficient in languages such as Rust and C/C++, and is capable of building end-to-end functional applications from a single prompt, including complex 3D simulations using Three.js.