Grok 4.5: xAI’s coding-first model lands in Cursor and the API

In mid-July 2026, Grok 4.5 shipped as SpaceXAI’s (xAI) strongest model yet for coding, agentic work, and broader knowledge tasks. It was trained jointly with Cursor and available the same day in the Cursor editor and the SpaceXAI API.

Primary sources: Introducing Grok 4.5 (SpaceXAI) and Cursor’s launch note. Consumer apps (grok.com, X, iOS, Android) followed about a week later.

This is not “another chat model with a fun persona.” The pitch is long-running tool use: investigate, call tools, recover from mistakes, and verify results across software engineering and adjacent desk work.

What is different about 4.5

Per Cursor and SpaceXAI:

• Mixture-of-experts model trained jointly; Cursor contributed a large slice of anonymized interaction data so the model sees how people and agents actually move through codebases and tools.
• Training mix stayed broader than Composer-class specialists: STEM, papers, and knowledge work, not only repo edits.
• Reinforcement learning on hard environments meant to stay difficult even for frontier models. That is the usual “agents that can fail and retry” story, with more emphasis on verification.

Cursor also noted new cybersecurity-oriented safeguards reflecting the model’s capabilities. Stronger agent models without policy friction are a liability for vendors; expect refusals and routing on sensitive offensive topics.

Composer 2.5 remains a separate, lighter weight class. Grok 4.5 is the bigger, broader model, not a rename of Composer.

Where you can use it

At launch (July 16):

• Cursor: desktop, web, iOS, CLI, and SDK; individual and team plans include significant first-party usage (launch week even doubled that allotment).
• SpaceXAI console / API: standard chat and completions-style access with an API key.
• Grok Build and various gateways (per SpaceXAI’s announcement packaging).

Consumer Grok surfaces got the model shortly after (announced July 22). If your team lives in Cursor already, you do not need a separate consumer account to evaluate it for engineering work.

Cursor’s published API-style pricing for the base model landed around $2 / $6 per million input/output tokens, with a faster variant at higher rates. Confirm current console pricing before you budget a migration. These numbers move.

Practical take for MSP and ops teams

We already use AI IDEs for PowerShell and bash drafts. Grok 4.5 sits in that same category of tool: useful when it shortens the path to a reviewed script, dangerous when someone pastes a generated change into production Graph or a firewall.

Worth trying:

• Multi-file refactors and “finish this agent task” loops inside a lab repo.
• Turning a messy ticket into a structured runbook draft you still edit.
• Long sessions where earlier models lost context or stopped verifying.

Still non-negotiable:

• No secrets, tenant dumps, or customer PII in prompts.
• Lab or staging first; -WhatIf / dry-run before live.
• Second pair of eyes on anything that modifies more than one object.

See also: AI IDEs for ops scripts. The workflow advice ages better than any single model name.

Bottom line

Grok 4.5 is xAI’s mid-2026 frontier push, co-trained with Cursor and aimed at agentic coding and desk work rather than meme answers. Try it where you already review AI output. Do not treat a launch blog as a reason to skip change control.