Now Under SpaceXAI, 'Grok-4.5' Debuts As An Opus-Class Model Focused On Coding, Agents, And Lower Costs

The large language models (LLMs) war isn't ending anytime soon. And smarter models keep coming, bringing new things tot he table.

During the competition that sparked following the rise of OpenAI's ChatGPT, the story of Grok began in late 2023 when xAI introduced its first model as a chatbot designed to answer a wide range of questions with real-time information drawn from the X platform and a deliberately less restrictive approach than many contemporaries.

Grok-1 arrived as a 314 billion parameter mixture-of-experts system and was later released with open weights. Grok-2 added image generation capabilities. Then, Grok-3 leveraged substantial new compute resources to improve performance across mathematics, coding, and instruction following, while introducing dedicated modes for deeper search and step-by-step reasoning.

Grok 4 arrived in July 2025 as the next major step.

It incorporated native tool use, real-time web and platform search, multimodal understanding, and a voice interface. A heavier multi-agent variant handled especially complex problems by running parallel lines of inquiry before synthesizing results. At the time it posted leading or near-leading results on several academic and reasoning benchmarks and became available through subscriptions and an API for developers and organizations.

Now, it's 'Grok-4.5.'

Released in July 8th, following xAI's rebrand to SpaceXAI, the model continues this progression with a deliberate specialization.

It is the first model in the line trained explicitly for coding and agentic workflows, developed in close coordination with the Cursor environment.

The training process emphasized large-scale codebases, multi-repository projects, extended tool interactions, and tasks that unfold over many steps or hours. Early demonstrations show it constructing functional applications, generating dashboards from financial data, producing engineering visualizations such as rocket assemblies, and managing project timelines with structured outputs.

Efficiency forms a central part of the update.

The model completes comparable work while using roughly half the output tokens of several leading alternatives and runs at approximately 80 tokens per second.

According to SpaceXAI in the announcement, the model has "twice greater token efficiency" than other leading models.

And if that carries through to real-world use cases, that efficiency would be a big advantage for SpaceXAI, since the cost of tokens has been a growing concern for AI consumers.

API pricing stands at $2 per million input tokens and $6 per million output tokens.

On software engineering benchmarks that measure resolution rates for real repository issues and terminal-based tasks, Grok 4.5 records scores that sit close to or ahead of current frontier entries from other laboratories, with particularly large gains visible once token consumption is taken into account.

Internal evaluations described by Elon Musk place its overall capability roughly on the level of advanced Claude Opus versions around the 4.7 mark while delivering markedly higher speed.

In a post on his X account which is now a subsidiary of SpaceXAI, Musk compared the Opus model, designed for intensive and complex tasks.

"Based on strong positive feedback from customers in our beta test program, @SpaceXAI will make Grok 4.5 available to the public tomorrow. It is an Opus-class model, but faster, more token-efficient and lower cost," he said.

It's worth noting that the current competition extends beyond building the smartest models or topping benchmark leaderboards. Equally important is developing systems that can scale efficiently and operate reliably in real-world environments.

Grok-4.5 appears to be taking that approach.