Skip to content

Model Card: Grok 4.7

Sep 23, 20261 min
TL;DRGrok 4.7 shipped 2026-09-21 as `grok-4.7`: a new, larger base model with longer RL training; 500K-token context window; API pricing unchanged at $2.00 input / $6.00 output per 1M tokens (rising to $4.00/$12.00 above 200K tokens). xAI's own Terminal-Bench 4.0 score jumps from 20.3% to 38.0%, but Artificial Analysis's independent re-test puts it at only 26% — well behind GPT-6 Astra (60%) and Claude Fable 5.1 (55%). It rolled out to all GitHub Copilot plans the same day. For agent builders, it's a 3-8x cheaper option for agentic coding, but complex autonomous terminal work still warrants independent benchmarking before you rely on it.

🌏 中文版

Model Info

FieldValue
Model IDgrok-4.7
VendorxAI (SpaceXAI)
Parameters~2.1 trillion (per third-party reporting; xAI has not published an exact figure on its spec page)
Context Window500,000 tokens
Input pricing (USD/1M tokens)$2.00 (< 200K prompt tokens) / $4.00 (≥ 200K)
Output pricing (USD/1M tokens)$6.00 (< 200K prompt tokens) / $12.00 (≥ 200K)
Open sourceNo
Release date2026-09-21
Official announcementxAI News: Introducing Grok 4.7
FamilyGrok 4.x (Grok 4.5 → Grok 4.6 → Grok 4.7, this release)

Highlights

  • Official Terminal-Bench 4.0 score jumps from Grok 4.6's 20.3% to 38.0% (+17.7pp) — the largest gain of any benchmark in this release
  • EEBench (electrical-engineering tasks) hits 64.0%, beating both the prior generation (53.0%, +11.0pp) and every competitor xAI cited: GPT-5.6 Sol Max (39.4%) and Claude Fable 5.1 Max (56.4%)
  • Built on an entirely new, larger base model, with a longer reinforcement-learning run weighted toward harder, longer-horizon tasks — xAI frames the training goal as "finishing the work," not just producing a plausible-looking answer
  • A rebuilt safeguard stack: on HackerBench v0.3 (malicious cyber tasks), the pass-through rate for dangerous dual-use prompts drops to just 3.3%; on LatchBio's biosafety benchmark it posts the top score of 62.4%

Benchmarks

BenchmarkGrok 4.7Previous gen (Grok 4.6)Best competitor
CursorBench 4.046.3%40.4%Claude Fable 5.1 Max 51.8%
DeepSWE v1.1 (high)71.0%65.2%GPT-5.6 Sol Max 72.7%
Terminal-Bench 4.0 (xAI self-reported)38.0%20.3%Claude Fable 5.1 Max 57.9%
EEBench64.0%53.0%Best in field (GPT-5.6 Sol Max 39.4%)
Artificial Analysis Intelligence Index (independent)46No directly comparable prior scoreClaude Fable 5.1 / GPT-6 tied at 53
Terminal-Bench 4.0 (Artificial Analysis re-test, standardized conditions)26%GPT-6 Astra 60%, Claude Fable 5.1 55%

⚠️ Except for the two Artificial Analysis rows, every figure above is self-reported by xAI or the respective vendor, each run at its own highest reasoning-effort setting (high/max) — these are not apples-to-apples comparisons and await further independent replication.

vs. Previous Generation / Competitors

Against Grok 4.6, Grok 4.7's biggest gains land in agentic coding: CursorBench 4.0 (+5.9pp), DeepSWE v1.1 (+5.8pp), and — most dramatically — a near-doubling on Terminal-Bench 4.0 (20.3% → 38.0%). xAI attributes this to swapping in a larger new base model and up-weighting training data toward difficult, hours-long tasks.

The competitive picture is more mixed. On xAI's own numbers, Grok 4.7 beats GPT-5.6 Sol Max on CursorBench 4.0 and EEBench, and edges out Claude Fable 5.1 Max on DeepSWE v1.1 — but its self-reported Terminal-Bench 4.0 score still trails Fable 5.1 Max by nearly 20 points. More notably, when Artificial Analysis re-ran Terminal-Bench 4.0 under standardized conditions, Grok 4.7 scored just 26% — 12 points below xAI's own 38.0% figure, and far behind GPT-6 Astra's 60% and Claude Fable 5.1's 55%. On that same firm's composite Intelligence Index, Grok 4.7 scores 46, seven points behind the co-leaders Claude Fable 5.1 and GPT-6 (53 each), placing it in the middle of the frontier pack.

Pricing is unchanged: $2/$6 (below 200K tokens), identical to Grok 4.5 and 4.6. That puts Grok 4.7's output pricing at a third of GPT-5.6 Sol's and an eighth of Claude Fable 5.1's — a cost advantage that compounds in agentic workloads with long reasoning chains and heavy output-token usage.

What This Means for Agent Development

Grok 4.7's training focus — sustaining long-horizon tasks and stronger self-verification, plus native understanding of xAI's own Grok Bot harness — directly benefits teams already building on the Grok ecosystem. But the 12-point gap between xAI's self-reported Terminal-Bench 4.0 score and Artificial Analysis's independent re-test suggests the official numbers may be optimistic for real-world autonomous terminal work.

  • If you're building cost-sensitive agentic coding products (e.g., a SaaS running coding agents at scale): the $2/$6 pricing, combined with scores that match or slightly beat GPT-5.6 Sol / Fable 5.1 on CursorBench and DeepSWE, makes Grok 4.7 worth evaluating as a drop-in replacement for some of your higher-cost calls — especially if you already use GitHub Copilot, Cursor, or Grok Build, where the switch is available day one
  • If you're building an agent for electrical engineering or a similar vertical knowledge-work domain: the top EEBench score is a signal worth A/B testing in that specific niche
  • Not a good fit: high-reliability, complex autonomous terminal work (multi-step shell commands, handling unexpected errors over long sessions) — Artificial Analysis's standardized re-test shows a real gap versus GPT-6 Astra and Claude Fable 5.1 here. Basing capacity planning on xAI's self-reported 38.0% may be overly optimistic; run your own task set through independent evaluation before betting heavily on this capability

Today's Takeaway

Grok 4.7's Terminal-Bench 4.0 score is 38.0% by xAI's own testing, but only 26% under Artificial Analysis's standardized independent re-test — a 12-point gap on the exact same benchmark name. That's a reminder worth internalizing: a vendor's self-reported benchmark score and a third party's re-test under controlled conditions are not the same measurement, especially for agentic and terminal-operation benchmarks, where differences in tooling versions, timeout settings, and reasoning-effort levels can swing the result. When independent re-test data exists, don't stop at the vendor's own table.

References