xAI has moved Grok 4.5 into private beta, restricting initial access to SpaceX and Tesla, where engineers are running the model against real production workloads before any outside release. Elon Musk posted the announcement on X on June 28, 2026. No public benchmark suite, pricing tier, or general availability date accompanied the disclosure.
The model sits on a V9 foundation with 1.5 trillion parameters. xAI supplemented training with data from Cursor, the AI-assisted coding editor, positioning Grok 4.5 as a model where software engineering tasks are a primary design target rather than an afterthought. Reinforcement learning is still running, meaning the weights that SpaceX and Tesla are testing today are not the weights that will ship publicly.
Musk described early internal evaluations as placing Grok 4.5 near or above Anthropic’s Claude Opus. That claim deserves structural scrutiny: xAI conducted the evaluations, xAI owns the infrastructure where they ran, and xAI has a financial interest in the result. No third-party organization has published benchmark numbers. No academic lab has reproduced the evaluation. The claim is Musk’s, stated by Musk, measured by Musk’s company, on a timeline set by Musk. Buyers and builders should treat it as a directional signal at most until independent replication exists.
The choice to beta-test inside SpaceX and Tesla is worth noting as a structural move, not just a marketing beat. Both companies generate dense, domain-specific technical data: aerospace engineering documentation, vehicle diagnostics, manufacturing telemetry. Internal deployment gives xAI a feedback loop that most labs have to simulate. Google can test Gemini on search logs; Meta can test Llama on social-graph data; xAI now has a direct channel into two verticals where reasoning under constraint, not just code generation, is the operative task. Whether that feedback actually accelerates the public version depends on how tightly xAI closes the loop between user sessions and training runs.
xAI’s model cadence has accelerated substantially in 2026. Grok 3 shipped in February with a heavy emphasis on reasoning; Grok 3.5 arrived in April targeting coding specifically. The step to Grok 4.5 follows a roughly six-week interval, which is consistent with a lab running RL post-training as a continuous process rather than a discrete release event. The naming suggests xAI is treating this as an intermediate release rather than a generational leap, though 1.5 trillion parameters at the foundation layer is not a small increment.
The Cursor data inclusion is the detail most likely to matter for developers evaluating xAI as a coding partner. Cursor’s dataset represents actual developer workflows: file context management, multi-file edits, test-driven iteration, refactoring patterns. Adding it in supplemental training is a direct response to criticism that earlier Grok versions underperformed on realistic software engineering tasks relative to Claude and GPT-4o. Whether supplemental Cursor data moves the needle on tasks outside the Cursor distribution is unanswered.
For teams currently comparing coding models, the practical situation is this: Grok 4.5 is not available to evaluate, Musk’s internal eval claims are unverified, and the RL run has not completed. The model that eventually ships publicly will differ from what SpaceX and Tesla are using today. Watch for third-party benchmark coverage from Chatbot Arena or METR before committing any infrastructure decisions to xAI’s coding stack.
Posted by Elon Musk on X on June 28, 2026.
