SpaceXAI shipped Grok 4.7 on Sep 21 at the same $2/$6 pricing as 4.6, with CursorBench 4.0 at 46.3% (vs 40.4%) and native Grok Bot harness training. Artificial Analysis independently scores Intelligence Index 46 (+2) and Coding Agent Index 56 with Grok Build (+9), at higher token use. GitHub Copilot is rolling the model out Pro→Enterprise. Creator “INCREDIBLE” packaging is not a universal ranking.

Alex Finn — Grok 4.7 inside Grok Bot is INCREDIBLE

What the video shows

Required embed for this package: Alex Finn — “Grok 4.7 inside Grok Bot is INCREDIBLE” (YouTube ajuOF7vclCM), published September 21, 2026. House note: this is creator packaging, not a SpaceXAI company upload and not an Artificial Analysis report. Use it as a viewer’s walkthrough of Grok 4.7 inside Grok Bot workflows — and keep “INCREDIBLE” out of AISN’s factual claim column. Prefer this embed with hedges over NONE because it matches the locked Grok Bot Tuesday mix and lands on the same news day as the primary post. Do not let the title upgrade a same-price model launch into a verified win over Opus or Astra.

What’s new

AI Shift News already covered Grok Bot Galaxy as a live product-education event, the August plan expansion / Berman use-case lane, and the X connector scaffolding. Those posts are soft-re-cover bans for today’s lead. What is new is a dated model release: Grok 4.7 on September 21, with SpaceXAI’s own CursorBench delta, unchanged list pricing, Fast-tier tradeoffs, and an independent AA pass that both raises coding-agent scores and documents the token bill.

That combination matters for house voice. Labs routinely ship “best ever” language. The scarce journalistic job is to keep the same-dollar upgrade, the independent index moves, the rows where competitors still lead, and the creator packaging in the same paragraph.

Evidence

SpaceXAI primary — Introducing Grok 4.7 (Sep 21). SpaceXAI describes Grok 4.7 as its most capable model for coding and knowledge work: a new, larger base versus Grok 4.6; a longer reinforcement-learning run on a harder mix weighted toward tasks that take many hours; better self-verification and longer-context management; and native training to understand the Grok Bot harness for conversational and knowledge-work tasks. On CursorBench 4.0, which SpaceXAI says stresses longer-running coding tasks, Grok 4.7 is posted at 46.3% versus Grok 4.6 High at 40.4%. List pricing starts at $2 per million input tokens and $6 per million output tokens — the same as 4.6. A Fast variant is served at roughly twice the output speed at twice the price. Availability is stated for Cursor, Grok Build, the Grok API, third-party coding harnesses, and model routers / cloud platforms. SpaceXAI also posts a comparison table that includes GPT-5.6 Sol Max and Fable 5.1 Max; that table is useful precisely because it does not show Grok winning every column — Fable 5.1 Max is ahead on CursorBench 4.0 (51.8%) and Terminal-Bench 4.0 (57.9% vs Grok’s 38.0%). House fence: quote the 46.3% upgrade honestly; do not delete the competitor rows that keep “everywhere” off the table.

Artificial Analysis independent (Sep 21). AA scores Grok 4.7 (xhigh) at 46 on the Artificial Analysis Intelligence Index (+2 over Grok 4.6), bringing SpaceXAI into AA’s “top 4 labs” framing for that index. On AA-Briefcase, Grok 4.7 lands at 1657 Elo (+111 vs 4.6 high), just behind Claude Opus 5 and Claude Fable 5.1. With Grok Build, Coding Agent Index rises to 56 (+9 from 47), ranking 4th among models in their native harnesses — behind Fable 5.1, GPT-6 Astra, and Claude Opus 5. Component lifts include DeepSWE v1.1 (~65%→73%), Terminal-Bench 4.0 (~18%→33%), and SWE-Atlas-QnA (~58%→63%). AA also flags the cost of those gains: Grok 4.7 (xhigh) uses roughly 81k output tokens per Intelligence Index task versus about 38k for Grok 4.6 (xhigh) / ~36k for 4.6 high — and ~27k for GPT-6 Astra (max). Context window stays 500k; pricing matches 4.6 with cache hits discounted to $0.50 per MTok. That is the independent fence AISN wants: real score movement, real token bill, no universal crown.

GitHub Copilot changelog (Sep 21). GitHub says Grok 4.7 is rolling out for Copilot Pro, Pro+, Max, Business, and Enterprise, selectable in the model picker across VS Code, Visual Studio, Copilot CLI, cloud agent, Copilot app, JetBrains, Xcode, and Eclipse. Rollout is gradual. Enterprise/Business admins manage access via Copilot model policy; under default enablement, new models arrive unless admins disabled the global default or this model. Billing is at provider list pricing under usage-based billing. This is a distribution receipt, not a quality proof.

What is still missing (confirmed fence). No AISN-run bake-off in this package; no claim that Grok Bot seat conversion moved on the launch day; no erasure of AA’s “behind Fable/Astra/Opus” Coding Agent ranking; no treatment of creator YouTube as an eval lab.

Pricing continuity is part of the story, not a footnote. Holding $2/$6 while posting a CursorBench lift is SpaceXAI’s commercial claim that buyers can upgrade without a list-price jump — while the Fast tier makes the speed/cost trade explicit. Artificial Analysis’s cache-hit note ($0.50 per MTok on hits) and unchanged 500k context window keep the procurement sheet honest: what changed is capability and token burn, not the headline sticker or the window size. Readers comparing Cursor, Grok Build, and Copilot should treat “same price” as list-price continuity, then verify billable tokens in their own harness.

What this does not prove

  • It does not prove Grok 4.7 beats Opus, Astra, or Fable everywhere. SpaceXAI’s own table and AA’s native-harness ranking both keep competitors ahead on key columns.
  • It does not prove “INCREDIBLE” as an independent score. That is Alex Finn packaging.
  • It does not prove better scores are free. AA documents roughly 2× token use vs Grok 4.6 on Intelligence Index tasks.
  • It does not prove Grok Bot Galaxy outcomes, Aug 26 plan ROI, or X-connector product readiness. Those are separate AISN files — do not soft-re-lead them.
  • It does not prove Copilot availability equals enterprise endorsement of every Grok 4.7 claim. Changelog = distribution + admin policy.

Why it matters

For practical readers buying coding agents, the scarce signal is often same-dollar movement plus an independent desk that is willing to publish both the score and the token bill. Grok 4.7 clears that bar on September 21: SpaceXAI held list price while posting a CursorBench lift, AA independently moved Coding Agent and Briefcase numbers, and GitHub started a Copilot rollout. That is useful procurement context — if you refuse to confuse it with a crowns ceremony.

It also matters for the Grok Bot Tuesday franchise specifically. Native Bot-harness training is the product-thread that connects a model card to the assistant readers actually open. Cover the model upgrade; do not pretend the assistant’s prior Galaxy/demo packages are today’s news.

What to watch next

  1. AA / third-party follow-ups — whether Coding Agent Index holds after more labs re-run harnesses.
  2. Token economics in real Cursor / Grok Build sessions — whether ~2× token use shows up as bill shock or acceptable quality spend.
  3. Copilot enablement reality — when Enterprise admins actually see and allow the model.
  4. SpaceXAI Fast variant adoption — who pays 2× for 2× speed.
  5. Amplifier hygiene — outlets that upgrade “46.3%” into “beats Opus everywhere” without AA’s ranking caveats.

Bottom Line

Grok 4.7 (Sep 21) is a same-price frontier coding upgrade: CursorBench 4.0 46.3% vs 4.6’s 40.4% at $2/$6, native Grok Bot harness training, AA Intelligence Index 46 (+2) and Coding Agent Index 56 (+9) with Grok Build, plus a gradual GitHub Copilot rollout — and a documented jump to ~81k tokens/task. Treat Alex Finn’s “INCREDIBLE” title as packaging. It is not proof Grok 4.7 beats Opus/Astra/Fable on every bench. Prefer xAI + AA + GitHub; keep claim and proof in separate columns.