xAI Ships Grok 4.7 With Larger Model and Better Coding at Unchanged $2/$6 Pricing

Grok 4.7 is xAI's latest reasoning model, optimized for coding and multi-hour agentic knowledge work. The base model grows to roughly 2.1 trillion parameters — about 40% larger than Grok 4.6 — trained with a longer reinforcement-learning run and, per xAI, incorporating SpaceX engineering datasets for complex technical work. Crucially, pricing holds steady at $2/$6 per million tokens, keeping xAI in the same low-cost bracket as OpenAI's new Sol tier.

Benchmarks show real gains: CursorBench 4.0 rose to 46.3% from 40.4%, and DeepSWE v1.1 hit 71.0%. A new safeguard stack shows a 3.3% risky-prompt pass-through rate, and xAI shipped Grok Voice Transcribe 2.0 alongside at a 6.8% word error rate across 19 languages. The model rolled out instantly to GitHub Copilot, giving it immediate distribution.

The skeptical counterpoint, flagged by VentureBeat and others, is token economics: Grok 4.7's multi-hour reasoning tasks consume large volumes of tokens, so the affordable per-token headline can mask high per-task cost and erode real-world ROI. Developer reaction was comparatively muted, with the community treating 4.7 as an incremental step in a fast release cadence rather than a leap — a pattern reinforced by the fact that a Grok 4.6 Bedrock listing was still fresh when 4.7 arrived. Positioned against Opus 5.5 and GPT-6, Grok 4.7's pitch is coding parity at a stable price, but it will need to prove the token overhead doesn't cancel the sticker savings.