Cursor + Grok 4.5: How to Enable It and When to Use It

Grok 4.5 is available inside Cursor on all plans, including free, and shows up as a selectable model in the model picker. xAI and Cursor co-trained the release, so it lands as a first-class option rather than a bring-your-own-key experiment.

This guide walks through enabling it, matching it to the right Cursor mode, and knowing when Claude Sonnet should stay your default. We stick to what xAI and Cursor have published and flag every number that requires you to check the vendor page.

For teams evaluating a model switch across an engineering org, we cover the workflow trade-offs first and the sticker price second. Both matter, but the mode-fit call decides whether the switch sticks.

How to enable Grok 4.5 in Cursor

Open Cursor Settings, go to the Models tab, and toggle Grok 4.5 on in the list of available models. It appears natively for every Cursor plan, including free, so you do not need to paste an xAI API key or configure a custom base URL to try it.

If Grok 4.5 does not appear, update Cursor to the latest version. Restart the editor after enabling the model so the picker in the chat sidebar refreshes.

One known gotcha: if you have an OpenAI custom API key or a custom base URL configured under Settings > Models > OpenAI, community reports suggest disabling those while using Grok 4.5, then re-enabling them when you switch back to a custom-provider model.

Once enabled, Grok 4.5 shows up in the model dropdown at the top of the chat panel, in inline edits (Cmd+K), and in Composer. Pick it per-request rather than as a global default until you have run it against your own repo.

Cursor mode-fit matrix for Grok 4.5

Cursor exposes four main ways to invoke a model: Ask (chat), Edit (Cmd+K inline), Agent (autonomous multi-step), and Composer (multi-file). Grok 4.5 is not equally strong in all four, and picking the wrong mode is the most common reason a new model feels worse than your default.

Use Grok 4.5 in Agent and Composer first. xAI positioned the release around multi-step software engineering and agentic tool use, and Cursor's own launch post frames it the same way. That is where the joint training shows up.

Use Grok 4.5 in Ask when you need a longer reasoning pass over a wide slice of code — xAI publishes a 500K context window for the model, which is roughly two to four times what most default coding models carry. Verify the current context limit on docs.x.ai before you plan a workflow around it.

Keep your existing default for tight inline Edits (Cmd+K on a selection). Inline edits reward latency and predictability more than raw capability, and switching models mid-flow adds friction without a clear payoff.

Prompt patterns that work in Cursor with Grok 4.5

Give it the whole problem, not a slice. Grok 4.5's large context window means you can drop in a failing test, the module under test, the shared types file, and the README section that describes intended behavior — all in one Agent turn. Small, keyhole prompts underuse the model.

Ask for a plan before code on any change touching more than one file. In Composer, prompt with: state the goal, list the files involved, propose the sequence of edits, then produce them. This mirrors how xAI trained the model for multi-step engineering tasks.

For Agent runs, name the exit condition. 'Run until the test suite passes and no lint errors remain' beats 'fix the bug' because it gives the model a stopping rule and something to check against. Grok 4.5 will keep pulling tools if you leave the goal open.

When you switch back to a different model inside the same conversation, start a new chat. Cursor keeps context per thread, and mixing model outputs in one thread makes it hard to tell which model wrote what.

Cost and latency versus Cursor's Claude Sonnet default

xAI publishes Grok 4.5 pricing at $2 per million input tokens and $6 per million output tokens, served at roughly 80 tokens per second, per the Cursor launch post and third-party coverage. Confirm the current rate on the xAI console before you budget against it — model prices move.

Anthropic publishes Claude Sonnet pricing on its own site, and the effective per-request cost inside Cursor depends on whether you are on a Pro plan (included usage) or usage-based billing. Compare the numbers on the vendor pages the day you decide.

Latency-wise, Grok 4.5 at 80 tokens per second is competitive for chat and Composer but noticeable on long generations. If your team lives inside inline Cmd+K edits all day, a small per-edit latency delta compounds fast — measure before you switch defaults org-wide.

xAI's pitch is 'highest intelligence per unit of time and cost' driven partly by fewer output tokens per task. That claim is worth testing on your own three or four representative tickets rather than taking on faith.

When to keep Claude Sonnet as your Cursor default

Keep Sonnet as the default for teams whose work is mostly inline edits, small refactors, and short chat exchanges. That is where a familiar model with predictable behavior beats a marginal capability upgrade.

Keep Sonnet where your codebase is heavily regulated or has strict output-review workflows. Introducing a new model changes the shape of diffs your reviewers see, and the first two weeks of adoption always add review overhead. Do that intentionally, not on a whim.

Keep Sonnet if you are in the EU. Coverage at launch flagged that Grok 4.5 was not yet available in the EU — check the xAI availability page before rolling it out to European engineers.

Switch to Grok 4.5 as the default only after a two-week bake-off on real tickets, measured against your own quality bar. Vendor benchmark scores are a starting point, not the decision.

Grok 4.5 inside a Cursor workflow audit

In our own work running the /keyword-gap and /mindmap-pass routines across dozens of AI-content sites in our portfolio, the pattern we see with a new model launch is the same every time: three weeks of hype, then a real answer about where it belongs in the stack. Grok 4.5 in Cursor is at week two.

At Layer3Labs we treat model selection the way we treat any procurement call for a client — start with the jobs the model actually needs to do, then map each job to the cheapest model that clears the quality bar. Cursor makes that easy because the picker is per-request.

Our default recommendation for engineering teams evaluating Grok 4.5 today is to enable it, route Agent and Composer runs to it for a sprint, and keep Cmd+K on your incumbent. Revisit after real reviews of real diffs.

This is the same pattern we use inside the AI Workflow Audit — pick one workflow, measure the current tool, test the challenger, decide with data.

Where to verify anything in this guide

Every number that moves — price, context window, rate limit, availability by region — should be checked on the primary source before you budget or plan around it. Vendors update these pages without release notes.

For Grok 4.5 pricing and rate limits, check the xAI console and docs.x.ai. For Cursor-side behavior (which modes support the model, plan availability, any pricing markup inside Cursor), check the Cursor blog and the Cursor Models settings page.

For benchmark claims, do not rely on any single blog post — including this one. Cross-check against the model card xAI publishes and, where possible, an independent evaluation.