GitHub Copilot Adds Grok 4.6 Across CLI, IDE, and Cloud

GitHub didn't wait long. Two days after SpaceXAI shipped Grok 4.6 on August 12, 2026, GitHub's own changelog confirmed the model landed in Copilot's model picker — not as a staged rollout across a few surfaces, but across eight at once: VS Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains IDEs, Xcode, and Eclipse. Same day, same model, same picker entry, everywhere.

That "everywhere, at once" detail is the actual story for practitioners. Model additions to Copilot aren't new — Kimi K2.7 Code became the first open-weight model in the picker back in July, and Grok 4.5 preceded 4.6 by about two weeks. What's changed is that GitHub is now treating cross-surface parity as a launch requirement instead of an afterthought.

TL;DR

What actually changed: eight surfaces, one day

Read GitHub's own model-hosting reference and the picker is already crowded — OpenAI's GPT-5.6 line, Anthropic's Claude family up through Opus 5 and Fable 5, Google's Gemini 3.7 Flash, Microsoft's own MAI-Code models, and the open-weight Kimi entries. Grok 4.6 doesn't expand that list with a new provider — xAI was already there via Grok 4.5. What it demonstrates is GitHub's willingness to push a brand-new model (two days old) to every client surface at once, rather than the historical pattern of IDE-first, CLI-later, cloud-agent-eventually.

That pattern matters because Copilot's surfaces have genuinely different jobs. VS Code and JetBrains are inline-completion and chat contexts. The Copilot CLI is for terminal-driven, agentic multi-step work — closer in spirit to Claude Code or Cursor's agent mode than to autocomplete. The cloud agent runs unattended, PR-driving workflows. A model tuned for "longer-horizon tasks requiring sustained reasoning and tool use" — GitHub's own phrase for Grok 4.6 in the changelog — is specifically suited to the CLI and cloud-agent surfaces, less so to inline completions where latency dominates.

The operational change: pick once, not per tool

Before this rollout pattern, a developer standardizing a team on a non-default model had to check availability surface by surface — was it in the VS Code picker yet, did the CLI flag exist, had the cloud agent config caught up. Grok Code Fast 1's original 2025 rollout, and even Grok 4.5's July 2026 launch, staggered across surfaces over days or weeks.

Shipping to all eight on day two removes that reconciliation step. A team picking Grok 4.6 as their default now gets the same model behavior whether a developer is typing in VS Code, running gh copilot in a terminal, or letting the cloud agent draft a PR overnight — no re-authentication, no "wait, is it live in JetBrains yet" checking. That consistency is worth more to a team's actual velocity than the marginal benchmark delta between Grok 4.6 and whatever else is in the picker, because model-switching friction across tools is a real, measured productivity tax — not a hypothetical one.

Enterprise and Business plans are the exception worth flagging: the Grok 4.6 policy defaults to off, same as it did for Kimi K2.7 Code and for Grok 4.5 before it. An admin has to explicitly enable it in Copilot settings before any developer on those plans sees it in their picker, regardless of surface.

Where Grok 4.6 actually sits against Claude and GPT-5.6

GitHub's changelog is deliberately silent on comparisons — it says nothing about how Grok 4.6 stacks up against the Claude or OpenAI models already in the same picker. That's a real gap for a developer trying to decide what to actually select. SpaceXAI's own launch evals fill part of it: Grok 4.6 ties GPT-5.6 Sol at 61 on the AA Intelligence Index, and leads on GDPVal-AA v2, AA-Briefcase, and Harvey LAB — knowledge-work-flavored benchmarks. It does not sweep coding evals; Fable 5 Max still leads several of those, and a separate comparison against Gemini 3.7 Flash and Sonnet 5 found Grok 4.6 absent from Google's own launch charts entirely.

Price is where Grok 4.6 separates cleanly from the rest of the picker. At $2 input / $6 output per million tokens — unchanged from Grok 4.5 — it undercuts Claude Opus-class and GPT-5.6 Sol pricing by a wide margin. Perplexity's own developer changelog verification put the real gap against Fable 5 specifically at 80-88% cheaper per token, not the 60% figure that circulated after launch — worth knowing before you assume the marketing number is the real one.

A practical picking rule

Copilot's model picker itself is still a manual choice per chat session or CLI invocation — there's no built-in Cursor-Router-style auto-routing inside Copilot as of this rollout. That's a meaningful difference worth knowing if your team has gotten used to automatic model selection elsewhere.

Enterprise rollout mechanics

For Business and Enterprise admins, the enablement path mirrors every prior third-party model GitHub has added:

1. Confirm your plan is Business or Enterprise (Pro/Pro+/Max users see it automatically once their account rolls in — no admin action needed).
2. Go to Copilot organization settings and locate the Grok 4.6 policy toggle — it ships off by default.
3. Review data-governance implications before enabling; Grok 4.6 requests route through xAI's own infrastructure, same as Grok 4.5 did, not GitHub's Azure stack the way Kimi K2.7 Code does.
4. Expect gradual visibility even after enabling — GitHub's changelog explicitly notes the rollout is staged per account, so "enabled" and "visible to every developer today" aren't the same moment.

The bigger pattern: Copilot's picker as a live leaderboard

Zoom out and Copilot's model picker has quietly become one of the most consequential distribution channels in the industry — a new model's presence there, same-day across every surface, is now treated as a launch milestone worth its own changelog post, on par with the model release itself. That's a shift from 2025, when Copilot model additions lagged flagship launches by weeks. xAI, Moonshot, and Microsoft's own MAI team are all now shipping into Copilot within days of releasing anything, because the picker is where a meaningful share of enterprise developer usage actually happens.

For a developer, the practical takeaway isn't "Grok 4.6 is now the best model in Copilot" — it demonstrably isn't across every benchmark. It's that the cost of trying it is now near zero: no separate signup, no API key, no per-surface reconfiguration. Toggle the picker, run your normal CLI or IDE workflow, and compare the diff quality and latency against whatever you were using before switching a whole team over.

Version specs, pricing, and surface availability reflect GitHub's changelog and SpaceXAI's published evals as of August 16, 2026 — Copilot's model lineup and enterprise policy defaults change frequently; check GitHub's live model-hosting docs before standardizing a team on any single model.

Related reading

• Kimi K2.7 in GitHub Copilot: First Open-Weight Model
• Grok 4.6 Launch: Official Evals, Same $2/$6, Cursor Access
• Perplexity Adds Grok 4.6: Does 60% Cheaper Really Match Fable 5?
• Fable 5 vs Grok 4.6 vs GPT-5.6 Sol vs Qwen3-8-Max
• Gemini 3.7 Flash vs Grok 4.6 vs Sonnet 5 vs GPT-5.6
• Claude Code vs Cursor vs GitHub Copilot
• Cursor Router: Automatic Model Selection
• GitHub's Grok 4.6 changelog · SpaceXAI's Grok 4.6 in Copilot announcement