Grok 4.7 Release: Same Price, Longer Horizons

xAI released Grok 4.7 on September 21, 2026. Official surfaces brand the company as SpaceXAI. For SEO and prior Grok coverage, we keep the xAI / Grok naming readers already search. The release is a direct upgrade to Grok 4.6 at the same list price and the same 500K context class.

The thesis is simple. Larger base. Longer RL on multi-hour tasks. Stronger self-verification. Native Grok Bot harness understanding. A new safeguard stack. You are not buying a new price tier. You are buying more reliable progress on long-horizon coding and agent work at $2 / $6 under 200k.

Headline self-reported numbers: 46.3% on CursorBench 4.0 (xhigh), 71.0% on DeepSWE v1.1 (high), 38.0% on Terminal-Bench 4.0 (xhigh, Grok Build), 66.0% on EEBench (xhigh, model card). None of these are LLM Stats verified yet.

At a Glance

• Release date: September 21, 2026. Generally available on named day-one surfaces.
• Model ID: grok-4.7 on the Grok API.
• Pricing: $2 / $0.50 cached / $6 per 1M under 200k prompt tokens. $4 / $1 / $12 at or above 200k. The higher tier bills the whole request. Matches Grok 4.6.
• Context: 500,000 tokens. Pretraining cutoff June 2026. Supplemental training through August 2026.
• Modalities: Text + image in, text out.
• Reasoning: reasoning.effort low / medium / high (default) / xhigh.
• Where: Cursor, Grok Build, Grok API, third-party harnesses, routers, and clouds (per the announcement).
• Rate limits (docs): 150 rps, 50M tokens/min. Regions: us-east-1, us-west-2, us-central-1.

What's New in 4.7

None of these change the sticker price or the context class. Together they push Grok further into long, less-attended work.

Larger base and longer multi-hour RL

Per the launch post, Grok 4.7 uses a new, larger base than 4.6, then a longer reinforcement learning run on a harder mix weighted toward problems that take many hours. The model card adds that supplemental training ran longer than for 4.6, with agentic RL across knowledge work, general coding, and purpose-built environments.

Self-verification and longer context management

xAI's product claim is fewer steps and fewer output tokens on hard tasks, with more careful self-checking on long trajectories. CursorBench 4.0 is the cleanest public signal for that story: longer horizon than CursorBench 3.x, so scores are not comparable across versions.

Native Grok Bot harness understanding

The announcement says Grok 4.7 was trained to natively understand the Grok Bot harness. Read that as better fit for conversational and knowledge-work loops that sit on that stack, not as a free win in every third-party agent.

Effort rails

Docs expose four effort levels. Default is high. Several buyer-facing benches in the card are reported at xhigh. Treat the ladder as a cost and latency dial, then measure your own token-per-task.

Benchmarks

All scores below are self-reported by xAI in the announcement and the Grok 4.7 model card (September 21, 2026). Where those two sources disagree, we use the card. Not LLM Stats verified. The launch chart compares 4.7 to 4.6 only. Secondary domains below are absolute card scores.

Coding and long-horizon agents

CursorBench 4.0 is the buyer number for multi-hour IDE agents. Grok 4.7 hits 46.3% at xhigh (43.9% at high). The news table puts Grok 4.6 at 40.4% at high, so the +5.9 point headline is a cross-effort comparison, not a same-rung gain. Still the price-performance claim xAI leads with at the same $2 / $6 list price. Version 4.0 is not comparable to CursorBench 3.2 from the Grok 4.6 post.

DeepSWE v1.1 moves 65.2% → 71.0% at high. That is steady repository-level progress, not a cliff. SWE-Marathon v1.1 jumps harder: 31.9% → 46.0% at high. If your workload looks like long agent streaks rather than single-issue patches, Marathon is the more informative delta.

Terminal-Bench 4.0 is the largest relative move in the coding set: 20.3% at high → 38.0% at xhigh. Caveat: Grok's score uses the Grok Build harness. Compare harness to harness before you treat peer tables as apples to apples.

FrontierSWE V2 rises 25.3% → 29.0% at xhigh on Proximal's Proximus harness. The card is explicit: the headline is average partial-credit reward, not task resolution. Useful as a ultra-long-horizon signal. Wrong as a "pass rate."

Engineering acceleration

EEBench and CADGenBench sit next to coding as engineering acceleration, not as another SWE suite. EEBenchuses the card's 66.0% at xhigh (news comparison table lists 64.0%). Vs Grok 4.6 at xhigh that is 60.0% → 66.0%. CADGenBench is 44.4% at highon Grok Build: the card's generation split for CAD construction. Read EEBench as electrical engineering agent work; CADGen as whether the model can produce usable CAD construction steps, not as a coding pass rate.

Legal and clinical (launch table)

Harvey remains low in absolute terms even after the gain. HealthBench Pro's +8.2 is material if clinical reasoning is on your evaluation sheet. Neither replaces domain review with a human expert. The bio/lab rows below sit next to HealthBench as adjacent capability probes, not as a clinical product claim.

Cyber, bio / lab, and CAD (absolute)

These card rows do not have clean 4.6 deltas in the launch table, so they stay out of the hero chart. Grouped below by job. Cyber numbers are unrestricted capability scores from the card. Bio/lab and WMDP rows are capability or knowledge signals, not refusal metrics.

On cyber: CyberGym at 80.3% (high, Mean Reproduced %, unrestricted / without standard safeguards, Grok Build) is the strongest absolute signal in the set. CVE-Bench at 36.6% (xhigh primary reward; card also notes 37.7% at high) and CathedralBench at 29% (xhigh, hard-subset accuracy, third-party red-team cyber) are harder probes. Treat them as capability measurements under the card's stated conditions.

On bio / lab: LatchBio Capabilities v1.0 at 44.5% (xhigh, equal-weight mean of 11 benches) is the composite. Single-suite highs include VCT 63.0%, LAB-Bench Practical 76.8%, ProtocolQA Open-Ended 70.4%, and BixBench 88.4% (all high except LatchBio at xhigh). WMDP-Bio / Chem / Cyber land at 88.1% / 84.9% / 88.1% (high). Those are dual-use knowledge MCQs. Useful as knowledge coverage. Not a biosafety policy statement.

Knowledge-work Elo

AA Briefcase v1.1 moves 1546 → 1657 on the news table. GDPval is listed at 1695 Elo (xhigh) in announcement materials. Elo is not percent. Treat both as directional knowledge-work signals, not as a coding verdict.

Pricing & Availability

The 200k cliff is still the cost story people miss. Cross 200k prompt tokens and every token in that request bills at the higher rate. List price matches 4.6, so migration is not a budget renegotiation. See docs.x.ai/docs/models/grok-4.7 for the live model page.

Availability per the announcement: Cursor, Grok Build, the Grok API, plus third-party coding harnesses, model routers, and cloud platforms. We only quote public xAI list pricing here.

When to Use / Migration from 4.6

Swap the model id to grok-4.7. Same price class. Same context class. Three checks before you flip agent traffic at scale.

1. Decide whether long-horizon coding is the job

Prefer 4.7 when Cursor-style multi-hour coding, terminal agents, or marathon-style streaks matter. CursorBench 4.0 and Terminal-Bench 4.0 are the clearest public deltas. If your traffic is short chat with light tool use, re-measure before you assume the upgrade pays for itself in quality alone.

2. Re-measure token-per-task at high and xhigh

Default effort is high. Several vendor headlines are xhigh. Output can rise when the model stays in the job longer. Do not assume 4.6 token budgets still hold.

3. Watch the 200k cliff; note the fast variant once

Long prompts still flip the whole request to $4 / $1 / $12. The announcement mentions a fast variant at 2× price and 2× output speed. Docs currently expose grok-4.7 only. Plan latency lanes from what the API actually lists in your account, not from a marketing aside.

• Good fit: teams already on Grok 4.6 who want more persistence on hard coding and agent loops without changing price.
• Also fit: Cursor and Grok Build users who want the new default model on those surfaces.
• Hold / verify first: regulated legal or clinical workflows. Absolute Harvey and HealthBench scores still need human review.

Caveats

• Self-reported table. Every number in this post is from xAI materials. Not LLM Stats verified.
• Harness coupling.Terminal-Bench 4.0, EEBench, CADGenBench, and CyberGym for Grok use Grok Build. FrontierSWE V2 uses Proximal's Proximus harness and reports partial-credit reward, not resolution. CyberGym is unrestricted Mean Reproduced %.
• EEBench disagreement. Model card: 66.0% at xhigh. News comparison table: 64.0%. We cite the card and keep the footnote.
• CursorBench version break. 4.0 is not comparable to CursorBench 3.2 from the Grok 4.6 writeup.
• Fast variant API surface. Mentioned in the launch post. No separate public API id on the model docs page at publish time.
• Not for unattended high-stakes decisions. The model card states Grok 4.7 is not intended for autonomous high-stakes use in medicine, law, finance, or safety-critical systems without human oversight.

Outlook

Grok 4.7 is the same commercial shape as 4.6 with a harder training diet. The benches that move most are the ones that punish short attention: CursorBench 4.0, Terminal-Bench 4.0, SWE-Marathon, HealthBench Pro. The platform story is continuity. Same $2 / $6 opening price. Same 500K window. Same model-id swap path.

For buyers, the decision is whether long-horizon coding and agent persistence are the bottleneck. If they are, 4.7 is the obvious next Grok step. If they are not, treat the release as a free quality bump on an unchanged bill, then verify on your own evals.

Primary sources: the xAI launch post, the model docs page, and the Grok 4.7 model card (September 21, 2026).