Grok 4.7: xAI's Frontier Coding and Knowledge Work Model with New Safeguard Stack and Longer RL Training

Executive Summary

On September 21, 2026, xAI (SpaceXAI) released Grok 4.7, its most capable model for coding, agentic tasks, and knowledge work. Built on a new, larger base model than Grok 4.6, Grok 4.7 was trained with a longer reinforcement learning run on a harder mix of tasks, specifically weighted toward problems that take many hours to complete. The model is better at verifying its own work, managing longer context, and natively understands the Grok Bot harness for conversational tasks and general knowledge work.

At $2 per million input tokens and $6 per million output tokens — the same price as Grok 4.6 — Grok 4.7 delivers significant improvements across its benchmark suite: 46.3% on CursorBench 4.0 (up from 40.4%), 71.0% on DeepSWE v1.1 (up from 65.2%), 64.0% on EEBench (up from 53.0%), and 19.6% on Harvey Legal Agent (up from 15.8%). The model supports 500K context, configurable reasoning effort (low, medium, high, xhigh), and encrypted reasoning content on the Responses API.

The release also introduces an entirely new safeguard stack that represents xAI's strongest safety calibration to date. Grok 4.7 tops LatchBio's biosafety benchmark at 62.4% and achieves the highest safety on HackerBench v0.3, allowing only 3.3% of risky dual-use prompts through while rarely blocking legitimate security research. Select cybersecurity partners have been given invite-only access to Grok 4.7's red-team capabilities for defense research.

What makes Grok 4.7 notable in the October 2026 frontier landscape is not just the raw score improvements — it's the training methodology shift toward longer, harder tasks, the comprehensive safety overhaul, and the price-performance positioning that undercuts GPT-5.6 Sol and Fable 5.1 by 2×–5× while remaining competitive on key benchmarks.

---

1. What Is Grok 4.7?

1.1 The Training Methodology Shift

Grok 4.7 represents a fundamental change in how xAI trains its frontier models. Rather than simply scaling up parameters, the focus shifted to training duration and task difficulty:

"Grok 4.7 uses a new, larger base model compared to Grok 4.6. It was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete."

This is a departure from the typical pattern of releasing incrementally larger models. Instead, xAI invested in:

• Longer RL training runs — more iterations of reinforcement learning on harder problems
• Multi-hour task weighting — prioritizing tasks that require sustained reasoning over extended periods
• Self-verification — explicit training on checking one's own work
• Longer context management — better handling of extended conversations and codebases

1.2 Grok Bot Native Understanding

A unique capability of Grok 4.7 is its native understanding of the Grok Bot harness. Grok Bot, launched in August 2026, provides "durable AI teammates that work on a persistent cloud computer, with messaging, approvals, connectors, and routines." Grok 4.7 was specifically trained to work within this harness, making it better at:

• Conversational tasks — maintaining context across extended interactions
• General knowledge work — research, analysis, document creation
• Multi-step workflows — tasks requiring tool invocation and state management

This native integration gives Grok 4.7 an advantage in agent-based workflows compared to models that are merely compatible with agent frameworks.

1.3 Model Properties

Property	Value
Model name	grok-4.7
Context window	500,000 tokens
Knowledge cutoff	May 2026
Input modalities	Text, Image
Output modalities	Text
Output limit	No text output limit
Reasoning effort	Low, Medium, High (default), xhigh
Input price	$2.00 / 1M tokens
Output price	$6.00 / 1M tokens
Cached input	$0.50 / 1M tokens
Long-context pricing	$4/$1/$12 (input/cached/output) above 200K prompt tokens
APIs	Responses API, Chat Completions
Tools	Function calling, web search, X search, code execution
Regional endpoint	US endpoint available (10% premium)

---

2. Benchmark Performance: The Multi-Hour Task Advantage

2.1 xAI's Official Comparison

xAI published a comprehensive comparison table on the Grok 4.7 launch page, pitting 4.7 against Grok 4.6, GPT-5.6 Sol, and Fable 5.1 across seven benchmarks:

Benchmark	Grok 4.7 xHigh	Grok 4.6 High	GPT-5.6 Sol Max	Fable 5.1 Max
CursorBench 4.0 (software engineering)	46.3%	40.4%	41.7%	51.8%
DeepSWE v1.1 (software engineering)	71.0%	65.2%	72.7%	70.0%
EEBench (electrical engineering)	64.0%	53.0%	39.4%	56.4%
AA Briefcase v1.1 (multi-hour office work)	1,657	1,546	1,487	1,678
Terminal-Bench 4.0 (multi-hour terminal)	37.6%	20.3%	37.3%	57.9%
Harvey Legal Agent (legal work)	19.6%	15.8%	2.5%	6.7%
HealthBench Professional (clinical reasoning)	56.7%	48.5%	60.5%	62.1%

2.2 Where Grok 4.7 Leads

Grok 4.7's most dramatic improvements are in domains requiring sustained reasoning:

• EEBench: A 25.6-point jump from 53.0% to 64.0%, and a 24.6-point lead over GPT-5.6 Sol (39.4%). This is the largest margin of victory in the comparison.
• Harvey Legal Agent: A 144% improvement from 15.8% to 19.6%, and a 6.8× lead over GPT-5.6 Sol (2.5%). Legal reasoning appears to be a particular strength.
• CursorBench 4.0: A 14.6% improvement from 40.4% to 46.3%, surpassing GPT-5.6 Sol (41.7%) by 4.6 points.
• Terminal-Bench 4.0: An 85% improvement from 20.3% to 37.6%, matching GPT-5.6 Sol (37.3%).

2.3 Where Grok 4.7 Trails

The gaps reveal where the frontier race remains competitive:

• CursorBench 4.0: Fable 5.1 still leads at 51.8% (vs. 46.3%).
• DeepSWE v1.1: GPT-5.6 Sol leads at 72.7% (vs. 71.0%), though the gap is narrow.
• Terminal-Bench 4.0: Fable 5.1 leads at 57.9% (vs. 37.6%).
• HealthBench Professional: Fable 5.1 leads at 62.1% and GPT-5.6 Sol at 60.5% (vs. 56.7%).

2.4 Independent Benchmark Results

Beyond xAI's self-reported numbers, independent evaluators have published results:

Benchmark	Score	Source
AA Intelligence Index	46.5%	Artificial Analysis
AA-HLE (Humanity's Last Exam)	43.1%	Artificial Analysis
AA AutomationBench	65.6%	Artificial Analysis
AA Terminal-Bench 4.0	25.8%	Artificial Analysis
CWE-bench v1	68.0%	Collinear
AA-SciCode	57.4%	Artificial Analysis
FrontierSWE v2	29.5%	Proximal
AA-LCR (Long Context Reasoning)	76.7%	Artificial Analysis
AA Omniscience Index	32.0%	Artificial Analysis

The independent results generally confirm xAI's self-reported scores, with some variation due to different evaluation harnesses and settings. Notably, the AA Terminal-Bench 4.0 score of 25.8% is lower than xAI's self-reported 37.6%, highlighting the importance of evaluation methodology.

---

3. Pricing: Same Price, Better Performance

3.1 Token Pricing

Price per 1M tokens	Grok 4.7	Grok 4.6	GPT-5.6 Sol	Fable 5.1
Input	$2.00	$2.00	$4.00	$10.00
Cached input	$0.50	$0.50	$0.20	$0.20
Output	$6.00	$6.00	$20.00	$50.00

Grok 4.7 maintains the same pricing as Grok 4.6 while delivering significant performance improvements. This creates a free upgrade for existing users and positions Grok 4.7 as the most cost-effective frontier model for coding and knowledge work:

• 2× cheaper than GPT-5.6 Sol on input ($2 vs. $4) and 3.3× cheaper on output ($6 vs. $20)
• 5× cheaper than Fable 5.1 on input ($2 vs. $10) and 8.3× cheaper on output ($6 vs. $50)

3.2 Long-Context Pricing

For prompts exceeding 200K tokens, pricing doubles:

Price per 1M tokens (200K+ prompt)	Grok 4.7
Input	$4.00
Cached input	$1.00
Output	$12.00

Even at long-context rates, Grok 4.7 remains competitive — $4/M input is still half of GPT-5.6 Sol's standard input price.

3.3 Grok 4.7 Fast

A fast variant is available with 2× the output speed at 2× the price (or 1.5× for long-context requests). This variant is exclusive to Cursor and Grok Build and is not available on the public xAI API:

Variant	Speed	Price Multiplier	Availability
Grok 4.7	Standard	1×	xAI API, Cursor, Grok Build, routers
Grok 4.7 Fast	2× output	2× (1.5× long-context)	Cursor, Grok Build only

---

4. The New Safeguard Stack

4.1 A Complete Safety Overhaul

Grok 4.7 was built with an entirely new safeguard stack, representing xAI's strongest safety calibration to date:

"It is the strongest model we've tested on refusals and jailbreak resistance. In dual-use domains like cybersecurity and biological work, it leads on both utility for benign tasks and safe refusal on dangerous ones."

This is a significant shift from previous Grok releases, where safety was often secondary to capability. The new stack addresses three key areas:

4.2 Biosafety Leadership

On LatchBio's biosafety benchmark, Grok 4.7 achieves 62.4%, the leading score among frontier models. This benchmark tests models on their ability to refuse dangerous biological requests while maintaining utility for legitimate research.

4.3 Cybersecurity Balance

On HackerBench v0.3 — xAI's internal benchmark for risky and malicious cyber tasks — Grok 4.7 allows only 3.3% of risky dual-use prompts through while rarely blocking legitimate security research. This represents a careful balance between:

• Safety: Blocking dangerous requests (weapon development, exploit generation, etc.)
• Utility: Allowing legitimate security research (vulnerability analysis, defense strategies, etc.)

4.4 Red-Team Access

xAI has started giving select cybersecurity partners invite-only access to Grok 4.7's red-team capabilities for defense research. This mirrors Google's Fairwind Program approach (see Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30) but is more limited in scope, focusing on cybersecurity rather than broader frontier access.

---

5. Technical Features and API Capabilities

5.1 Configurable Reasoning Effort

Grok 4.7 supports four levels of reasoning effort:

Level	Description	Use Case
Low	Minimal reasoning, fast response	Simple queries, routine edits
Medium	Balanced reasoning and speed	Standard coding tasks
High (default)	Deep reasoning for complex tasks	Default for most workloads
xhigh	Maximum reasoning effort	Hardest problems, multi-hour tasks

The xhigh setting is particularly relevant for the multi-hour tasks that Grok 4.7 was specifically trained to handle.

5.2 Encrypted Reasoning on Responses API

On the Responses API, Grok 4.7 always returns reasoning.encrypted_content, even when the include parameter does not list it. This enables multi-turn conversations to maintain the model's reasoning without extra configuration:

import os
from xai_sdk import Client
from xai_sdk.chat import user

client = Client(api_key=os.getenv("XAI_API_KEY"))

chat = client.chat.create(model="grok-4.7")
chat.append(user("Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"))

response = chat.sample()
print(response.content)

The encrypted reasoning content should be passed back unchanged in the next request's input to maintain context continuity.

5.3 Prompt Caching Best Practices

xAI strongly recommends setting a prompt_cache_key (Responses API) or x-grok-conv-id header (Chat Completions) to route a conversation's requests to the same server, making cache hits reliable. Without this, users often pay the full input price on a cache-cold server.

5.4 Tools and Integrations

Grok 4.7 supports a comprehensive set of tools:

• Function calling: Custom tool definitions for external API integration
• Web search: Real-time web information retrieval
• X search: Search within the X (Twitter) platform
• Code execution: Sandbox-based code running and testing

---

6. Deployment and Access

6.1 Availability Channels

Channel	Status	Notes
xAI API	✅ Available	grok-4.7 model slug
US Regional Endpoint	✅ Available	us.api.x.ai/v1, 10% premium
Cursor	✅ Available	All plans, including Fast variant
Grok Build	✅ Available	Default model, free tier included
Model Gateways	✅ Available	OpenRouter, Vercel, Cloudflare
Cloud Platforms	✅ Available	Third-party providers

6.2 Grok Build Integration

Grok 4.7 is the default model of Grok Build, xAI's coding agent. Users can get started for free at x.ai/build, with the Fast variant available as a paid upgrade.

6.3 Integration with Agent Frameworks

The Grok Bot harness (launched August 2026) provides durable AI teammates with persistent cloud computers, messaging, approvals, connectors, and routines. Grok 4.7's native understanding of this harness gives it an advantage in agent-based workflows.

---

7. Connection to Prior Research

7.1 The October 2026 Frontier Landscape

Grok 4.7 arrives in the most active period of the frontier AI race:

• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05 — Qwen3.8-Omni-Flash with native omnimodal agentic workflows
• Pareto 26 10 Preview Blended Model Pareto Frontier Cost Performance Unbiased 2026 10 02 — Unbiased's Pareto 26.10 Preview with blended-model cost-performance
• Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30 — Google's Gemini 4 Argon with 1M output tokens
• Openai Devday 2026 Dots Gpt 6 1 Sol Ultrafast Decisions Api Private Intelligence 2026 09 29 — OpenAI DevDay with GPT-6.1 Sol
• Claude Sonnet 5 5 Faster Cheaper Coding Jump Cyber Safeguards Distillation Protection 2026 09 29 — Anthropic's Sonnet 5.5
• This article — Grok 4.7 with longer RL training and new safeguard stack

7.2 The Cost-Performance Continuation

Grok 4.7 continues the cost-performance narrative established by recent releases:

Model	Input ($/1M)	Output ($/1M)	DeepSWE v1.1	CursorBench 4.0
Grok 4.7	$2.00	$6.00	71.0%	46.3%
GPT-5.6 Sol	$4.00	$20.00	72.7%	41.7%
Fable 5.1	$10.00	$50.00	70.0%	51.8%
Pareto 26.10	$0.80	$3.20	69.9%	—
Gemini 4 Argon	$2.00	$10.00	77.9%	—

Grok 4.7 sits in a sweet spot: cheaper than Sol and Fable while remaining competitive on key benchmarks. It doesn't match Pareto's extreme cost-efficiency ($0.80/$3.20) but offers a more established deployment path with better tooling.

7.3 The Safety Pacing Signal

Grok 4.7's new safeguard stack — with biosafety leadership, cybersecurity balance, and red-team access — continues the safety pacing signal that began with OpenAI's Astra shelving (see Openai Devday 2026 Dots Gpt 6 1 Sol Ultrafast Decisions Api Private Intelligence 2026 09 29) and Google's Fairwind Program (see Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30).

---

8. Key Takeaways

1. Longer RL on harder tasks is the differentiator. The shift from simply scaling parameters to investing in longer reinforcement learning runs on multi-hour tasks represents a fundamental change in training methodology. The results — particularly on EEBench (+25.6 points) and Harvey Legal (+144%) — suggest this approach pays off.

2. Price-performance is the killer feature. At $2/$6 per million tokens — the same as Grok 4.6 — Grok 4.7 delivers a free upgrade that undercuts GPT-5.6 Sol by 2×–3.3× and Fable 5.1 by 5×–8.3× while remaining competitive on benchmarks.

3. The new safeguard stack is comprehensive. Biosafety leadership (62.4% on LatchBio), cybersecurity balance (3.3% dual-use pass rate), and red-team access represent xAI's most serious safety investment to date.

4. Grok Bot native understanding is unique. Training the model to natively understand the Grok Bot harness gives it an advantage in agent-based workflows that other models lack.

5. The 500K context window is competitive. While not matching Gemini 4 Argon's 1M output or Qwen3.8-Omni-Flash's 1M context, 500K is sufficient for most coding and knowledge work tasks.

6. Fable 5.1 still leads on CursorBench. At 51.8% vs. 46.3%, Fable 5.1 remains the top model for Cursor-based coding tasks, though Grok 4.7 is closing the gap.

7. Independent results confirm the trend. Artificial Analysis and other independent evaluators generally confirm xAI's self-reported scores, with some variation due to evaluation methodology.

8. The Fast variant fills a specific niche. Available only in Cursor and Grok Build, the 2× speed variant at 2× price targets developers who need faster iteration on coding tasks.

---

9. Future Directions

9.1 What to Watch

1. Grok 4.8 development: Musk has announced Grok 4.8 as a 2.5-trillion-parameter model built on a from-scratch C++ training stack. The timeline and capabilities remain uncertain.

2. Grok 5 roadmap: Musk stated that Grok 5 will target AGI, up from his October 2025 estimate of "10% and rising." The path from 4.7 to 5 will be critical to watch.

3. Independent benchmark replication: The gap between xAI's self-reported Terminal-Bench 4.0 score (37.6%) and Artificial Analysis's independent score (25.8%) highlights the need for standardized evaluation.

4. Grok Bot ecosystem growth: The native understanding of Grok Bot could become a significant advantage if the Bot ecosystem grows and attracts third-party integrations.

5. Enterprise adoption: The combination of low cost, strong coding performance, and comprehensive safety could make Grok 4.7 attractive for enterprise coding and knowledge work — if the tooling matures.

6. Competitor response: If Grok 4.7's multi-hour task training approach proves effective, expect other providers to adopt similar training strategies.

7. Red-team access expansion: The invite-only cybersecurity red-team access could expand to broader domains or become a standard offering.

8. Open-weight competition: The continued rise of open-weight models could pressure xAI's closed-model economics, though Grok 4.7's price-performance positioning makes it resilient.

---

10. References & Resources

Official Sources

• Grok 4.7 Announcement — Official xAI launch post with benchmarks and demos
• Grok 4.7 Documentation — API usage, specifications, and integration guides
• xAI Release Notes — Full release history including Grok 4.7 details
• xAI Pricing — Token rates and plan details
• Grok Build — Coding agent with Grok 4.7 as default model
• Grok Bot — Durable AI teammates with persistent cloud computers

Related Research in This Journal

• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05
• Pareto 26 10 Preview Blended Model Pareto Frontier Cost Performance Unbiased 2026 10 02
• Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30
• Openai Devday 2026 Dots Gpt 6 1 Sol Ultrafast Decisions Api Private Intelligence 2026 09 29
• Claude Sonnet 5 5 Faster Cheaper Coding Jump Cyber Safeguards Distillation Protection 2026 09 29

External References

• BenchLM: Grok 4.7 — Benchmark aggregation and comparison
• Artificial Analysis: Grok 4.7 — Independent benchmark scores
• OpenRouter: Grok 4.7 — API pricing and provider listing

---

This article was researched using official xAI sources (Grok 4.7 announcement page, developer documentation, release notes, and pricing page) supplemented by benchmark data from BenchLM and Artificial Analysis. All benchmark figures from xAI's launch post are cited directly with appropriate caveats about vendor-reported results. Independent scores from Artificial Analysis are noted where they differ from self-reported numbers.