Announcements

xAI Releases Grok 4.5: Benchmarks, Integrations, and Pricing

xAI has launched Grok 4.5, a new large language model optimized for software engineering and knowledge tasks. It is trained on NVIDIA GB300 GPUs.

AIDeveloper44 Team

July 8, 2026·4 min read

xAI Releases Grok 4.5: Benchmarks, Integrations, and Pricing

Grok 4.5 aims to improve token efficiency and reasoning speed for software engineering tasks.

TL;DR

Grok 4.5 is a new model from xAI focused on software engineering, agentic workflows, and general knowledge tasks.
The model was trained on a cluster of NVIDIA GB300 GPUs and features deep integration with Microsoft Office and the Cursor IDE.
It aims for high token efficiency, averaging 15,954 output tokens per SWE Bench Pro task, which the company states is 4.2 times fewer than Opus 4.8 (max).
Pricing is set at $2 per million input tokens and $6 per million output tokens, with EU availability delayed until mid-July 2026.

On July 8, 2026, xAI announced the release of Grok 4.5. Developed as the latest iteration of the company's language models, Grok 4.5 is specifically engineered for software development, agentic task automation, and general knowledge work. According to the release documentation, the model was trained alongside the coding environment Cursor, signaling a targeted focus on practical programming applications.

Training Infrastructure and Methodology

Grok 4.5 was trained using a computing cluster comprising tens of thousands of NVIDIA GB300 GPUs. The company outlined a training methodology that emphasized data curation over raw token volume. This process involved strict deduplication, quality scoring, and domain-focused data selection designed to maintain high coverage and signal within the training mixture.

Furthermore, xAI implemented a reinforcement learning (RL) training approach prioritizing "per-token intelligence." The RL training protocol covered hundreds of thousands of multi-step software engineering tasks, utilizing both automated and model-based grading systems. The underlying software stack is designed for asynchronous training, allowing agentic rollouts to execute for extended periods while the system updates across the distributed GPU network.

Benchmark Performance

As part of the release, xAI published performance metrics for Grok 4.5 across several specialized benchmarks, comparing its capabilities to current industry models such as Fable, GPT 5.5, Opus 4.8, and GLM 5.2. In the DeepSWE 1.0 benchmark (evaluating pass@1 rates), Grok 4.5 scored 62.0%. This placed it behind Fable (max) at 66.1% and GPT 5.5 (xhigh) at 64.31%, but ahead of Opus 4.8 (max) which scored 55.75%.

On DeepSWE 1.1, the model achieved a 53% score, trailing Fable (max) at 70%, GPT 5.5 (xhigh) at 67%, and Opus 4.8 (max) at 59%, while outperforming GLM 5.2 at 44%. In the Terminal Bench 2.1 evaluation, Grok 4.5 secured a score of 83.3%, closely following Fable (max) at 84.3% and GPT 5.5 (xhigh) at 83.4%. Finally, on SWE Bench Pro, Grok 4.5 demonstrated a 64.7% resolve rate, positioning it below Fable (max) at 80.4% and Opus 4.8 (max) at 69.2%, but higher than Opus 4.7 (max) at 64.3% and GPT 5.5 (xhigh) at 58.6%.

Diagram: Abstract representation of the Grok 4.5 asynchronous reinforcement learning pipeline.

Inference Speed and Token Efficiency

xAI reports that Grok 4.5 operates at an inference speed of 80 tokens per second (TPS). A primary claim in the release is the model's token efficiency during complex problem-solving. During the SWE Bench Pro evaluations, Grok 4.5 averaged 15,954 output tokens per task. The developer contrasts this with the performance of Opus 4.8 (max), which utilized an average of 67,020 output tokens on the same tasks. According to xAI, this equates to a 4.2x reduction in token usage for Grok 4.5, potentially lowering operational costs for end-users executing prolonged agentic workflows.

Office Workflows and Application Generation

In addition to raw coding capabilities (spanning languages like Rust and C/C++), Grok 4.5 is capable of generating complete applications from single prompts. One example provided by xAI demonstrates the model generating a functional, modern three.js simulation of the solar system complete with an adjustable time user interface.

Grok 4.5 has also been integrated as the default model within the Grok Build platform and features dedicated plugins for Microsoft Office applications. In Excel, the model can synthesize web research, build multi-sheet formulas, and leave digital notes for future reference. Within PowerPoint, Grok 4.5 can generate slides using native shapes for diagram construction, while its integration in Microsoft Word focuses on drafting and document formatting.

Pricing and Availability

Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens via the xAI API. The model is currently accessible in Grok Build, available on all tiers within the Cursor IDE, and accessible directly through the xAI console. xAI is offering limited-time free usage of Grok 4.5 via the Grok Build interface and a downloadable CLI tool.

Due to regional compliance factors, Grok 4.5 is not currently available within the European Union through any of xAI's products or the API console. The company anticipates launching the model in the EU by mid-July 2026.

References & Sources