Grok 4.7 Advances Agentic Coding with Extended RL on Multi-Hour Tasks

Grok 4.7 is SpaceXAI's frontier model built for coding, agentic tasks, and knowledge work.

SpaceXAI released Grok 4.7 on September 21, 2026, delivering a frontier model that emphasizes agentic coding gains through extended reinforcement learning on multi-hour tasks. The model comes with a 500,000 token context window that enables it to handle large volumes of information without losing coherence over time. Pricing has been held steady at $2 per million input tokens and $6 per million output tokens to encourage broad adoption. This combination of features makes Grok 4.7 particularly suited for users engaged in lengthy coding sessions or knowledge intensive projects where maintaining context and accuracy is paramount. The model is positioned as an evolution that addresses the need for AI systems that can operate with greater autonomy in professional environments.

The release builds upon the capabilities of Grok 4.6 by incorporating a larger base model and subjecting it to more rigorous training protocols. This allows for better performance in areas that require sustained effort and self-correction. Users can expect the model to manage longer interactions more effectively due to the increased context capacity. The focus on tasks that take many hours to complete during training translates to practical benefits in real applications where projects do not resolve in short bursts. SpaceXAI has made the model available across several platforms to maximize its utility for different types of users.

What background led to the development of Grok 4.7?

Development of Grok 4.7 was informed by the performance characteristics of its predecessor, Grok 4.6, which provided the baseline for measuring improvements. The company identified areas where previous models could benefit from additional training focused on endurance and verification. By extending the reinforcement learning phase, the model acquires skills necessary for navigating the challenges of complex, time-consuming problems. This strategic choice reflects an understanding that many valuable tasks in coding and knowledge work do not fit into brief interaction patterns. The result is a model that can contribute more substantially to ongoing projects without requiring frequent resets or interventions from the user.

The broader context of the AI industry shows increasing interest in models that can function as agents capable of handling extended workflows. SpaceXAI's approach with Grok 4.7 aligns with this trend by prioritizing training on harder tasks. This not only improves benchmark scores but also prepares the model for deployment in scenarios where reliability over time is essential. Stakeholders in the field are watching how these changes affect the competitive positioning of frontier models from various providers. The stable pricing is a notable aspect that could influence purchasing decisions among budget-conscious organizations.

What new features distinguish Grok 4.7 in agentic coding?

Grok 4.7 introduces enhanced self-verification mechanisms that enable it to review its own outputs for errors or inconsistencies during the course of a task. This is complemented by improved tool use that allows the model to call upon external functions or data sources as part of its problem solving process. The capability for long-running tasks means it can maintain progress on projects that span multiple hours, keeping track of previous steps and adjusting as needed. These features collectively support more sophisticated agentic behaviors where the model acts with a degree of independence in coding environments. The emphasis on these areas stems from the training methodology that weighted multi-hour problems heavily in the mix.

Benchmark results provide evidence of these improvements with Grok 4.7 scoring 46.3% on CursorBench 4.0. It also achieves 71.0% on DeepSWE v1.1 when high effort is applied. These numbers come from evaluations conducted by SpaceXAI and highlight the model's strengths in relevant domains. The scores reflect the benefits of the longer reinforcement learning run and the larger base model used in its construction. Users interested in agentic coding will find these metrics useful for comparing the model against alternatives in the market.

What are the technical specifics behind the Grok 4.7 training and specifications?

The technical foundation of Grok 4.7 includes a new, larger base model that serves as the starting point for further training. This base model was then fine tuned with a longer reinforcement learning run that utilized a harder mix of tasks. The tasks were weighted toward those that take many hours to complete, which helps the model develop the necessary persistence and strategic thinking for such challenges. As a result, the model shows better performance in verifying its own work and in managing the large context window of 500,000 tokens. The documentation provides details on how to configure reasoning levels to suit different needs.

Pricing details specify $2.00 per million input tokens and $6.00 per million output tokens, matching the structure from the previous model. This pricing is detailed in the developer documentation to allow for accurate cost estimation in various use cases. The reasoning options include low for faster responses, medium for standard use, high as the default, and xhigh for the most thorough processing. These options give users control over the balance between speed and depth depending on the nature of their tasks. The combination of these technical elements makes Grok 4.7 a versatile tool for frontier applications.

Key specifications and performance of Grok 4.7

Metric	Value	Attribution
Context Window	500,000 tokens	SpaceXAI Documentation
Input Price	$2 per million tokens	SpaceXAI Documentation
Output Price	$6 per million tokens	SpaceXAI Documentation
CursorBench 4.0 Score	46.3%	SpaceXAI Announcement
DeepSWE v1.1 Score	71.0%	SpaceXAI Announcement

How does Grok 4.7 affect the market and its stakeholders?

1. Available in Cursor for direct coding assistance
2. Accessible through Grok Build for custom builds
3. Provided via the Grok API for programmatic access

The market implications include increased options for developers seeking high performance without changes to their cost structures. The unchanged pricing allows for seamless transition from previous versions for those already invested in the ecosystem. Stakeholders such as software companies and individual programmers can leverage the model in Cursor to enhance their coding productivity through better agentic support. The Grok API opens doors for integration into larger systems where custom agent behaviors can be developed around the model's strengths in long context and self-verification.

Enterprise users may find particular value in the model's ability to handle knowledge work over extended periods, potentially reducing the time spent on oversight and correction. The availability in multiple platforms ensures that different segments of the market can access the technology in the format that best suits their workflows. This broad accessibility could accelerate the integration of frontier models into everyday professional tools. The overall effect is to provide a competitive offering that balances advanced capabilities with economic predictability.

What reactions have authorities expressed about the Grok 4.7 release?

SpaceXAI has described Grok 4.7 as its most capable model for coding and knowledge work in the company announcement. The description emphasizes the model's ability to work longer on difficult tasks and to check its own work more carefully. These claims are supported by the benchmark results and the details of the training process. The mention of best-calibrated safeguards to date indicates attention to safety and reliability considerations in the model's design.

Grok 4.7 is our most capable model for coding and knowledge work. It works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.SpaceXAI, Company announcement

What can be expected next in the development of frontier models like Grok 4.7?

Future developments may continue to refine the reinforcement learning techniques to tackle even more demanding task sets. The current context window size of 500,000 tokens may serve as a base for further expansions to support even larger scale operations. Continued emphasis on self-verification and tool use will likely drive the creation of more advanced agentic systems. The pricing approach may remain a key factor in how these models are adopted across the industry.

Observers will look for how the performance on benchmarks like CursorBench and DeepSWE evolves in subsequent releases. The integration options in Cursor, Grok Build, and the API provide a foundation for ecosystem growth around the model. As more users engage with these features, additional insights into practical applications will emerge to inform future iterations. The focus on long-running tasks positions the line of models to meet the needs of complex, real-world projects in coding and knowledge domains.