- xAI has launched Grok 4.5, trained on tens of thousands of Nvidia GB300 GPUs, targeting programming, Agents, and knowledge-based tasks.
- On Terminal Bench 2.1, Grok 4.5 scored 83.3%, nearly on par with GPT 5.5 at 83.4% and only slightly lower than Fable 5 at 84.3%.
- On DeepSWE 1.1, Grok 4.5 scored 53%, lower than GPT 5.5 at 67% and Fable 5 at 70% in real-world GitHub issue resolution capabilities.
- On SWE Bench Pro, Grok 4.5 scored 64.7%, surpassing GLM 5.2 at 62.1% but lower than Opus 4.8 at 69.2% and Fable 5 at 80.4%.
- xAI stated that the model was trained using a rigorous data filtering process, removing duplicate data and selecting sector-specific data to enhance quality.
- The reinforcement learning phase utilized hundreds of thousands of tasks, primarily in software engineering, featuring an automated grading system and an asynchronous training infrastructure for Agents running for multiple hours.
- Grok 4.5 is priced at 2 USD per million input tokens and 6 USD per million output tokens, the lowest among the high-performance model tier.
- Meanwhile, Opus 4.8 is priced at 5 USD and 25 USD, GPT 5.5 and GPT 5.6 are priced at 5 USD and 30 USD, while Fable 5 reaches up to 10 USD and 50 USD per million input and output tokens, respectively.
- xAI stated that Grok 4.5 uses approximately 4.2 times fewer tokens than Opus 4.8 on SWE Bench Pro tests and achieves a speed of about 80 tokens per second.
- xAI’s strategy is similar to Chinese companies like Zhipu and DeepSeek: narrowing the performance gap and then competing on low pricing.
- Grok 4.5 is already available on Grok Build, Cursor, and the xAI Console platform, while also supporting plugins for Word, PowerPoint, and Excel.
- The model is not yet released in the European Union, with a launch expected in mid-July, and is being developed in parallel with the Cursor editor following SpaceX’s acquisition of the company in an all-stock transaction valued at 60 billion USD.
- 📌 Conclusion: Grok 4.5 does not lead every benchmark leaderboard but exerts massive pressure thanks to its very low pricing strategy. With fees at only about 20% of Fable 5 and significantly lower than GPT 5.5, combined with lower token consumption, it sharply reduces operational costs. If real-world efficiency matches the announcements, Grok 4.5 could become an attractive option for enterprises needing to balance performance and cost, rather than chasing the model with the highest benchmark score.
VIET NAM CONSULTING AND MEASUREMENT JOINT STOCK COMPANY
© 2026 Vietmetric