Grok 4.5 Arrives in Copilot with 500K Context and Three Levels of Reasoning: Where It’s Worth Using
By Devy · Category: AI Dev Tools — General
On July 28th, GitHub began the rollout of Grok 4.5 —xAI’s latest reasoning model— within GitHub Copilot. Not announced for later, not on a waitlist: shipping, gradually, to the model picker in VS Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains, Xcode and Eclipse, for Pro, Pro+, Max, Business and Enterprise.
“Gradual” is the word to hold onto, and it’s the first of two reasons why today you might search for Grok 4.5 in your dropdown and not find it. We’ll get to the second one, which is the more interesting of the two. But first: what actually shipped, and where it’s worth spending tokens.
What You Get
Three specs matter here.
A context window of 500,000 tokens. That’s the headline number, and it’s real: the same window that Grok 4.5 exposes through xAI’s own API. In practice it means you can throw a large chunk of a repository, a long build log, or a complete incident thread at it without needing to chunk.
Text and image input. Screenshots of a broken UI, an architecture diagram, a photo of a whiteboard: everything counts as valid input alongside your prompt.
Three levels of reasoning effort: low, medium and high. This is the part that changes how you work, not just how much fits. You’re not picking a model: you’re picking how much thinking you’re going to pay for on each task.
What GitHub Actually Noticed
The changelog contains an observation worth more than the specs sheet. GitHub reports that Grok 4.5 “performed especially well dispatching tools in parallel and taking direct action”, and that this made it effective for terminal tasks in VS Code and Copilot CLI.
It’s a narrow and specific claim, and it’s useful. Dispatching tools in parallel is exactly the bottleneck of agentic work in the terminal: an agent that fires git status, npm ls and a grep across the repo in a single turn instead of three sequential ones closes a diagnostic loop in a fraction of real time. If your day includes a lot of “figure out why the build broke” at the CLI, this is the profile you want.
Watch the framing, though: this is GitHub’s qualitative reading of its own platform, not a published benchmark. Take it as a strong hint about where to aim the model, not as a measured result.
How to Choose the Reasoning Level
There’s no official guide on this, so here’s the practical heuristic that applies to reasoning models in general:
- Low — mechanical work where the answer is mostly retrieval or transformation. Renaming across files, writing a commit message, translating a YAML config to TOML, explaining what a function does.
- Medium — the default for real work. Writing a feature against an existing codebase, debugging something with a clear reproduction, reviewing a diff.
- High — reserve it. Architecture decisions, a bug whose root cause you genuinely can’t locate, anything where a wrong answer costs you an hour of rework.
The reason to be disciplined about this isn’t purity. It’s the bill, which is the section that follows.
Two Things That’ll Trip You Up
One: on Business and Enterprise, the policy comes disabled by default. GitHub ships Grok 4.5 with its model policy disabled, which means nobody on your organization’s plan sees it in the picker until an admin enables it. If you’re on a corporate seat and Grok 4.5 doesn’t appear in your dropdown, it’s much more likely to be this than the gradual rollout not having reached you yet. The fix is a conversation with whoever administers your Copilot org, not waiting: the setting lives in your organization’s Copilot policies.
It’s worth understanding this as a pattern rather than an annoyance. Since Copilot became a multi-vendor model marketplace, GitHub’s default posture toward new third-party models is opt-in, not opt-out: the organization decides which vendors receive their code and prompts. That’s the right default from a governance perspective, and it’s also why “is it available?” and “can I use it?” are now two different questions.
Two: billing to provider list pricing under usage-based billing. As of June 1st, 2026, Copilot no longer counts premium requests: it consumes GitHub AI Credits, deducted based on token usage (input, output and cached) at each model’s published API rates. Each paid plan includes credits equivalent to its price: the $10/month Pro includes $10 in credits, the $19/user Business includes $19, the $39/user Enterprise includes $39. Grok 4.5 is deducted from that pool at xAI’s list rates, currently around $2 per million input tokens and $6 per million output tokens.
Do the math once, because it clarifies a lot. A 500K window filled to the brim is approximately a dollar of input per request: a tenth of a Pro plan’s total monthly credits, before the model has written a single line back. And when credits run out, there’s no fallback experience: individual users buy more, and organizations either allow overage at standard rates or cap it.
None of this makes Grok 4.5 expensive by market standards — $2/$6 is in the accessible range for a frontier reasoning model—. But it does mean the 500K window is a capability to spend deliberately, not a default to fill. Point it at the problem that needs the full repository in context; don’t hand it the full repository for a rename.
How to Try It in Ten Minutes
- Check your plan tier. Pro, Pro+, Max, Business or Enterprise.
- Open the model picker in VS Code or run
copilotin the CLI and look for Grok 4.5. If you’re on Business or Enterprise and it doesn’t appear, it’s the policy: ask your admin to enable it in your organization’s Copilot model policies. - Give it the task it’s reportedly good at. Open Copilot CLI in a repo with a genuinely broken build and ask it to diagnose the failure and propose a fix. Watch whether it dispatches multiple commands per turn: that parallel dispatch is the behavior GitHub highlighted, and you’ll see it in the transcript.
- Run the same task with effort medium and with high, and compare. You’ll learn more about where it’s worth paying for high with an A/B on your own codebase than with any benchmark table.
- Check your credit consumption afterward in the billing view, while the session is still fresh in your memory. Doing this once calibrates your intuition for the rest of the month.
The Takeaway
Grok 4.5 in Copilot isn’t a new artifact to install: it’s one more entry in a picker that already has several. What makes it worth ten minutes is a specific fit: a 500K window plus a documented strength in parallel tool dispatch makes it a good candidate for agentic terminal work in Copilot CLI, which is precisely the workflow where sequential tool calls hurt the most.
What makes it worth reading the fine print is that “available” now means two different things. On a personal plan, it’s there when the rollout reaches you. On a corporate plan, it’s there when someone decides it should be, and it bills at list price against a pool of credits you can drain faster than you’d expect with 500K of context.
What about you — have you tried it on Copilot CLI yet, or does your picker still not show it? If you’re on a Business or Enterprise plan, let us know if you had to ask your admin to enable the policy.