The Daily Downlink

Last pass

commentary

Grok 4.7 shipped late and priced like the Chinese layer

Grok 4.7 shipped Monday, closing the countdown this column has been clocking since the window opened in early September. The surprise is not the model — mid-pack on the Artificial Analysis Intelligence Index at 46 against Claude Fable 5.1 and GPT-6 at 53 each, and a 26% on Terminal-Bench 4.0’s agentic-coding set that trails GPT-6 Astra’s 60% and Fable 5.1’s 55%. It’s the price: $2 per million input tokens, $6 per million output, at a 500k context — a Western flagship priced where the Chinese open-weight layer lives. The release the industry watched as a capability event shipped as a price event.

The countdown that missed twice closed as a price event

What happened. SpaceXAI (xAI) launched Grok 4.7 with the model card live at docs.x.ai: 500k context, configurable reasoning, knowledge cutoff May 2026, priced $2 per million input tokens and $6 per million output. Built on a larger base model with longer reinforcement learning and better self-verification, and available through the Grok API, Cursor, and Grok Build. This is the release Musk booked for early September, pushed through a second window, then delayed with a public note about RL reward shaping — the countdown-follow-up this column owes now has its model card to read.

Why it matters. The price is the event hiding inside the launch. A Western frontier lab rating its flagship at $2/$6 puts the closed lane’s per-token cost in the same band traders reserve for open and Chinese models — exactly the discount operators were told only the open layer would deliver. But price is not the edge where agents actually run: Grok 4.7’s 26% on Terminal-Bench 4.0 agentic coding sits under DeepSeek V4.1 Flash at 27% and well behind the 55–60% of the closed pair. Cheap tokens buy a mid-pack model; they don’t buy agent-grade reliability. Shop on the coding bench, not the price card.

What I’m watching

Whether $2/$6 bends the rest of the frontier before DevDay on September 29 and Anthropic’s reported low-price card (still an unverified Tuesday leak, no card up yet). One take: the frontier’s value lane just got a discount floor from the model that was watched as a benchmark story — the pricing answer, not the Elo answer, is what competitors have to match first.

Source: docs.x.ai, the-decoder.com