Grok 4.5 Pricing Puts Enterprise Coding Costs Under Pressure

SpaceXAI launched Grok 4.5 for coding, agentic tasks and knowledge work, making it available through the API and in Grok Build and Cursor. The company priced it at $2 per million input tokens and $6 per million output tokens, and said its generation speed is 80 tokens per second.

According to the company's published results, the model scored 62% on DeepSWE 1.0 and 83.3% on Terminal Bench 2.1, while its ranking against competing models varied by benchmark. The company says it was trained on tens of thousands of NVIDIA GB300 units, with reinforcement learning focused on hundreds of thousands of multi-step software engineering tasks. These figures come from the developer itself and should therefore be treated as reported results, not an independent evaluation.

The lens: Economic transition The release matters to the region not because of its position on a single benchmark, but because its coding capabilities come with a published and relatively low usage price. Companies in the Middle East and North Africa can test coding assistants on real work such as system maintenance and application development, then measure the cost of a completed task rather than relying on token prices or developer results. The scale of the training hardware also highlights another gap. Consuming models through an API is accessible, but producing a competing model still requires immense capital and computing resources. The economic choice between buying the service, customising an open model, or investing in local training is therefore as much a sovereign decision as a technical one.

Source: https://x.ai/news/grok-4-5