Elon Musk (@elonmusk) said in a post on X on June 28 that Grok 4.5 is in private beta at SpaceX and Tesla, putting xAI's next flagship model first inside his own operating companies rather than into a broad consumer or API launch.
Musk described Grok 4.5 as based on a "1.5T V9 foundation model" with "Cursor data added in supplemental training." He also said early evaluations show performance "close to, perhaps exceeding Opus," a reference to Anthropic's Claude Opus line, and said reinforcement learning is continuing to improve the model.
That is the whole announcement, and the missing pieces matter. xAI has not published a Grok 4.5 model card, benchmark table, pricing page, API availability notice, or safety report. Musk did not specify which Opus model he was comparing against, which evals produced the internal result, whether the comparison covered coding, agentic work, general reasoning, or all of the above, or whether the Cursor data came from a licensing arrangement, opt-in product data, generated traces, public tasks, or another source.
The Cursor reference is the sharpest part of the post. Cursor, the AI code editor made by Anysphere, says in its data-use policy that customer data is not used for training when Privacy Mode is enabled, and that it maintains zero data retention agreements with all providers. The same policy says that when Privacy Mode is off, Cursor may use and store codebase data, prompts, editor actions, code snippets, and other code-related data and actions to improve AI features and train models, and that prompts and limited telemetry may be shared with model providers when users explicitly select their models. Musk's post does not explain what subset of that universe, if any, fed Grok 4.5.
For Musk, the rollout follows the same founder-controlled pattern that has defined Grok since xAI's launch: build a model, put it inside companies where he controls the workflow, and use those environments as pressure tests before wider distribution. SpaceX and Tesla are not neutral beta customers. They give Grok exposure to software, hardware, manufacturing, autonomy, robotics, operations, and support problems that are difficult to reproduce in public benchmarks. They also let Musk blur the line between customer validation and internal deployment. The structure is easier to understand after SpaceX's move earlier this year. On Feb. 2, xAI said SpaceX had acquired xAI. That made the private beta at SpaceX less a third-party trial than an internal deployment across Musk's consolidated AI and aerospace stack. Tesla, meanwhile, gives Grok a different testing surface: engineering software, vehicle software, factory processes, service, and the company's long-running autonomy ambitions. Musk did not say where inside either company Grok 4.5 is being used.
The new beta also lands less than six weeks after xAI introduced Grok Build, its terminal-based coding agent for SuperGrok and X Premium Plus subscribers. Grok Build is explicitly aimed at professional software engineering, with plan review, diffs, repository conventions, plugins, hooks, skills, MCP servers, headless mode, and parallel subagents. If Grok 4.5's supplemental training really does include meaningful Cursor-derived coding data, the commercial target is not just chat performance. It is the coding-agent market where Cursor, Anthropic's Claude Code, OpenAI's Codex line, GitHub Copilot, and a growing set of terminal-native agents are competing for developer workflows.
xAI's public baseline is Grok 4, which the company launched on July 9, 2025. In that release, xAI said Grok 4 included native tool use, real-time search integration, and API access, and it introduced SuperGrok Heavy for access to Grok 4 Heavy. xAI framed Grok 4 around scaled reinforcement learning on Colossus, its large GPU cluster, and said it had expanded verifiable training data beyond primarily math and coding into more domains.
Since then, xAI has pushed the Grok line toward enterprise and agentic use cases. Its Grok 4.1 release in November 2025 described a silent rollout across grok.com, X, and mobile apps before public availability, a useful precedent for the current private-beta posture. Days later, xAI announced Grok 4.1 Fast and an Agent Tools API, giving developers server-side tools for browsing the web, searching X posts, executing code, and retrieving uploaded documents.
Anthropic remains the benchmark Musk chose to name because Claude Opus has been widely positioned around difficult coding and agentic work. Anthropic said Claude Opus 4.5, released Nov. 24, 2025, was available in its apps, API, and all three major cloud platforms, and it emphasized software engineering, agents, computer use, deep research, and spreadsheet work. Musk's comparison to "Opus" is therefore a claim about the highest-value category in frontier AI: models that can operate tools, reason over codebases, and complete long-horizon work with fewer human interventions.
There is no public evidence yet that Grok 4.5 meets that bar. The only disclosed metric is Musk's own characterization of early internal evals. That does not make the claim meaningless, but it does put it in the same category as most frontier-model pre-announcements: strategically important, not independently testable.
The timing is still notable. xAI's site says Colossus reached 200,000 H100 GPUs in a single interconnected cluster and describes a roadmap to 1 million GPUs. In January, xAI said it had raised a $20 billion Series E, naming Valor Equity Partners, StepStone Group, Fidelity Management & Research Company, Qatar Investment Authority, MGX, Baron Capital Group, NVIDIA, and Cisco Investments among participants, and said the financing would accelerate compute infrastructure and product deployment. Grok 4.5 is the next public sign of what that capital and compute are being pointed at.
For now, Grok 4.5 is not a product customers can buy. It is a private model trial inside Musk-controlled companies, with a coding-data clue, an Opus comparison, and no independent scoreboard. That is enough to show where xAI is aiming: not just a better chatbot, but a model that can compete for the engineering workflows where frontier AI companies now expect durable revenue.