SpaceXAI's Grok 4.6 Rolls Out in GitHub Copilot, Targeting Agentic and Terminal-Based Coding

Overview

Grok 4.6, the reasoning model SpaceXAI released earlier this month, is now rolling out inside GitHub Copilot, according to GitHub’s changelog. GitHub describes the model as “designed for agentic coding and complex multi-step workflows,” adding it to Copilot’s model picker across Visual Studio Code, Visual Studio, Copilot CLI, JetBrains, Xcode, and Eclipse.

The integration follows Grok 4.6’s public debut on August 12, when SiliconANGLE reported that SpaceXAI — the company formerly known as xAI, renamed after its acquisition by SpaceX — released the model saying it “can outperform Anthropic PBC’s Claude Fable 5 in some areas.”

What We Know

• Copilot rollout. GitHub says that in its own internal testing, “Grok 4.6 showed strong results across terminal-based coding tasks in Visual Studio Code and Copilot CLI,” and that it “performed especially well on longer-horizon tasks requiring sustained reasoning and tool use,” according to GitHub’s changelog.
• Availability and billing. The model is rolling out gradually to Copilot Pro, Pro+, Max, Business, and Enterprise plans, billed at provider list pricing under usage-based billing, GitHub says. For Enterprise and Business customers, administrators must turn on the Grok 4.6 policy in Copilot settings, since it is off by default.
• Release cadence. Grok 4.6 arrived “only a month after SpaceXAI released its previous flagship LLM,” SiliconANGLE reports — a reference to Grok 4.5, which The Machine Herald previously reported launched in early July at $2 per million input tokens.
• Training process. SpaceXAI says it spent more time training Grok 4.6 than its predecessor, using an AI-generated dataset intended to improve reasoning and giving the model access to “high-quality engineering data,” according to SiliconANGLE. The initial training run was followed by supervised fine-tuning (SFT) and then reinforcement learning; SpaceXAI used Grok 4.5 itself to help optimize Grok 4.6’s SFT phase, with that optimization work focused on science and programming tasks, the outlet reports.
• Benchmark score. SpaceXAI evaluated Grok 4.6 on the Artificial Analysis Intelligence Index, a composite of nine benchmarks spanning science, coding, and financial services. Grok 4.6 scored 61 — on par with OpenAI’s GPT-5.6 Sol and one point behind Claude Fable 5, according to SiliconANGLE.
• Pricing and access points. The standard version of Grok 4.6 costs $2 per million input tokens and $6 per million output tokens, with a faster edition priced at twice that rate. Outside of GitHub Copilot, the model is also available through Cursor — the coding platform SpaceXAI acquired for $60 billion — and through SpaceXAI’s own Grok Build tool, SiliconANGLE reports.

What We Don’t Know

• GitHub’s changelog entry refers to the model provider as “xAI” rather than “SpaceXAI,” and does not address the rebrand; it is unclear whether this reflects lag in GitHub’s own copy or a distinction GitHub is drawing.
• Neither source specifies parameter count, context window size, or an exact rollout completion date for Copilot users beyond “gradual.”
• SiliconANGLE’s coverage of Grok 4.6’s performance against Claude Fable 5 on eight additional benchmarks (beyond the Artificial Analysis Intelligence Index) does not clearly specify which model — Grok 4.6 or its predecessor — the reported results apply to, so those figures are omitted here.

Analysis

The Copilot integration lands Grok 4.6 in front of a much larger developer audience than SpaceXAI’s own channels alone, and it follows a pattern GitHub has repeated through August: adding successive frontier coding models — including Microsoft’s own MAI-Code-1.1-Flash, as The Machine Herald previously reported — to Copilot’s model picker rather than betting on a single provider. For SpaceXAI, the placement extends Grok 4.6’s reach beyond Cursor and Grok Build into the IDEs and CLIs where much of Copilot’s existing user base already works, at a moment when the company is positioning the model as roughly competitive with GPT-5.6 Sol and Claude Fable 5 on composite benchmarks rather than clearly ahead of either.