What are the strengths of this model?

Terminal-Bench 4.0 more than doubles from 20.3% to 38.0%, covering the area where Grok 4.6 was weakest 46.3% on CursorBench 4.0 at $2/$6, closing most of the gap to Fable 5.1's $10/$50 while costing a fifth Best-calibrated safeguards to date per xAI: LatchBio biosafety 62.4%, only 3.3% of high-risk dual-use prompts slip through

What are the weaknesses of this model?

Still trails Claude on headline coding and office benches - Fable 5.1 tops CursorBench at 51.8% and Terminal-Bench at 57.9% Same sticker as Grok 4.6, so no per-token efficiency gain; the 2x Fast variant is the only speed-up, at double cost Native Grok Bot harness understanding ties the model to xAI's agent runtime rather than open tool ecosystems

What are the best use cases?

Long autonomous coding sessions where Terminal-Bench gains matter High-volume code review and debugging at the CursorBench tier Regulated or security-sensitive assists needing biosafety and dual-use gating

xAIProprietary

Grok 4.7

Name: Grok 4.7
Author: xAI

Compare this model

xAI's coding-and-knowledge-work flagship, released September 21, 2026. It keeps Grok 4.6's $2/$6 per-million pricing while extending reinforcement learning onto harder multi-hour tasks, improving self-verification, long-context handling and native understanding of the Grok Bot agent harness. The launch led with safety: a new safeguard stack that xAI says tops LatchBio biosafety at 62.4% and passes only 3.3% of high-risk dual-use prompts on HackerBench v0.3. On benchmarks it closes the gap to Claude Fable 5.1 on coding without reaching it - CursorBench 4.0 46.3% vs 51.8%, Terminal-Bench 4.0 38.0% vs 57.9% - but posts a real agentic leap over its predecessor, nearly doubling Terminal-Bench.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

500K

License

Proprietary

Release Date

2026-09-21

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

Billing Mode: standard

Strengths

•Terminal-Bench 4.0 more than doubles from 20.3% to 38.0%, covering the area where Grok 4.6 was weakest
•46.3% on CursorBench 4.0 at $2/$6, closing most of the gap to Fable 5.1's $10/$50 while costing a fifth
•Best-calibrated safeguards to date per xAI: LatchBio biosafety 62.4%, only 3.3% of high-risk dual-use prompts slip through

Weaknesses

•Still trails Claude on headline coding and office benches - Fable 5.1 tops CursorBench at 51.8% and Terminal-Bench at 57.9%
•Same sticker as Grok 4.6, so no per-token efficiency gain; the 2x Fast variant is the only speed-up, at double cost
•Native Grok Bot harness understanding ties the model to xAI's agent runtime rather than open tool ecosystems

Use Cases

•Long autonomous coding sessions where Terminal-Bench gains matter
•High-volume code review and debugging at the CursorBench tier
•Regulated or security-sensitive assists needing biosafety and dual-use gating

Deep Analysis

Context Window

500K tokens

Text and image input, text output, no output-length cap

API Price

$2 / $6 per 1M

Input/cached-input/output below 200K tokens; $4/$12 above; Fast variant doubles both

CursorBench 4.0

46.3%

vs Grok 4.6 40.4%, Claude Fable 5.1 (max) 51.8% (xAI-reported)

Terminal-Bench 4.0

38.0%

up from Grok 4.6's 20.3%; Claude Fable 5.1 57.9% (xAI-reported)

EEBench

64.0%

beats Claude Fable 5.1's 56.4% (xAI-reported)

LatchBio biosafety

62.4%

New safeguard stack; only 3.3% of high-risk dual-use prompts pass on HackerBench v0.3

Strengths

・Agentic leap: Terminal-Bench 4.0 more than doubles from 20.3% to 38.0%, the area where Grok 4.6 was weakest.
・Coding at a low price: 46.3% on CursorBench 4.0 at $2/$6 undercuts Claude Fable 5.1's $10/$50 while closing most of the gap.
・Best-calibrated safeguards to date per xAI, with a LatchBio biosafety score of 62.4% and only 3.3% of risky dual-use prompts slipping through.

Weaknesses

・Still trails Claude on the headline coding and office benches: Fable 5.1 maxes CursorBench at 51.8% and Terminal-Bench at 57.9%.
・Same sticker as Grok 4.6, so no efficiency gain per token; the 2x Fast variant is the only way to go faster, at double cost.
・Native Grok Bot harness understanding ties the model to xAI's agent runtime rather than open tool ecosystems.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Fable 5.1	—	71.0% (DeepSWE v1.1 high)	—	$10/$50 per 1M
GPT-5.6 Sol	—	—	94.6% (GPQA Diamond)	$5/$30 per 1M
Grok 4.6	—	65.2% (DeepSWE v1.1 high)	—	$2/$6 per 1M

Overview

Grok 4.7 is xAI's coding-and-knowledge-work flagship, released September 21, 2026. It holds Grok 4.6's $2/$6 pricing while extending reinforcement learning onto harder, multi-hour tasks, improving self-verification, long-context handling, and native understanding of the Grok Bot agent harness. The launch led with safety: xAI says its new safeguard stack tops LatchBio biosafety at 62.4% and passes only 3.3% of high-risk dual-use prompts on HackerBench v0.3. On benchmarks it closes the gap to Claude Fable 5.1 on coding without reaching it — CursorBench 4.0 46.3% vs 51.8%, Terminal-Bench 4.0 38.0% vs 57.9% — but posts a real agentic leap over its own predecessor. It ships in Cursor, Grok Build, the Grok API, and third-party harnesses, with a 500K-token context and a 2x-speed Fast variant at double price.

Benchmarks & Performance

Detailed Comparison

### Head-to-head with the models it actually competes with | Feature | Grok 4.7 | Claude Fable 5.1 | GPT-5.6 Sol | Grok 4.6 | | :--- | :--- | :--- | :--- | :--- | | Role | Coding / knowledge work | Absolute ceiling | OpenAI frontier tier | Prior xAI flagship | | Price (in/out per 1M) | $2 / $6 | $10 / $50 | $5 / $30 | $2 / $6 | | Context | 500K tokens | 1M tokens | 1.05M tokens | 500K tokens | | CursorBench 4.0 | 46.3% | 51.8% (max) | 41.7% | 40.4% | | Terminal-Bench 4.0 | 38.0% | 57.9% | 39.4% | 20.3% | | Native agent harness | Grok Bot | — | Codex | Grok Bot | The honest read: Grok 4.7 is a 4.6-class price with a much better agentic brain, but Fable 5.1 still owns the top of the coding and office charts. Pay for 4.7 when the job is long autonomous coding; pay for Fable when a single best result matters.

Community Feedback

Practitioner reaction centered on the safety framing. xAI led the launch post with its safeguard stack rather than raw scores, which some read as a bid for enterprise trust; others noted the benchmark gap to Fable 5.1 is real and that $2/$6 is unchanged from 4.6, so the win is agentic tool use, not price. The Fast variant drew interest from high-volume coding shops that already live in Cursor. A recurring complaint: native Grok Bot harness understanding is convenient inside xAI's stack but a lock-in outside it.

Use Cases

### Where Grok 4.7 earns its keep **1. Long autonomous coding sessions.** Terminal-Bench 4.0 jumping to 38.0% means it can actually drive tool-using agents instead of faking them. At $2/$6 this is the cheapest credible option for multi-step repo work. **2. Code review and debugging.** 46.3% on CursorBench 4.0 at $2/$6 undercuts Fable 5.1's $10/$50 for the same class of task, so it pays for itself on volume. **3. Regulated or security-sensitive assists.** The new biosafety and dual-use gating (LatchBio 62.4%, 3.3% pass-through) is the selling point here, not the raw score — defense partners get invite-only red-team access.

Latest News

**Launch (September 21, 2026).** xAI released Grok 4.7 as its most capable coding and knowledge-work model, a larger base than 4.6 trained with a longer RL run on multi-hour tasks. Pricing stays $2/$6 per 1M tokens (cached input $0.50; $4/$12 above 200K), with a 2x-speed Fast variant at double price. **Safety:** new safeguard stack, LatchBio biosafety 62.4%, HackerBench v0.3 only 3.3% risky dual-use pass-through, invite-only cyber red-team access for defense partners. **Availability:** Cursor, Grok Build, Grok API, third-party coding harnesses, model routers, cloud platforms. Native Grok Bot harness understanding is built in.

Sources

Analysis generated: 2026-09-24