xAI ships Grok 4.7 on 2.1T-parameter base, rolling out instantly to GitHub Copilot

Grok 4.7 arrives just over a month after Grok 4.6 and is tuned squarely for coding autonomy, professional knowledge work, and multi-hour reasoning, backed by a new safeguard stack. The base model grew to 2.1 trillion parameters, a 40% jump, which xAI credits for gains in self-verification and long-context handling.

The benchmark story is concrete: 46.3% on CursorBench 4.0 (up from 40.4%) and 71.0% on DeepSWE v1.1 at high effort, numbers developers on Reddit and HN praised as genuinely competitive for the price point. xAI kept token pricing unchanged at $2/$6 per million, and paired the release with Grok Voice Transcribe 2.0 (released September 18), which cuts word error rates from 20.6% to 6.8% across 19 languages at $0.10/hour batch and $0.20/hour streaming.

Distribution is a key differentiator: Grok 4.7 landed instantly in GitHub Copilot and supports model-router setups, lowering the switching cost for developers already inside those workflows. That easy-adoption path, plus flat pricing amid a week of aggressive cuts from OpenAI and Anthropic, keeps xAI cost-competitive.

The skeptical take, flagged by VentureBeat and developers alike, is token consumption: Grok's coding gains can come with high token usage on long multi-hour tasks, which threatens real-world ROI even at attractive per-token rates. In other words, the sticker price is competitive but the effective bill on agentic workloads is the number to watch. Grok 4.6 also shipped to Amazon Bedrock this week, giving enterprise teams a managed path to the prior generation while 4.7 rolls out.