Grok 4.7: Better Coding, Stronger Agents, Zero Price Increase

On September 21, 2026, xAI—operating in some contexts under the SpaceXAI banner following recent corporate alignments—officially launched Grok 4.7, its latest flagship large language model. Positioned explicitly for coding, agentic workflows, and professional knowledge work, the release marks a measured but meaningful step forward from Grok 4.6. The company emphasizes that the new model was built on a larger base architecture and subjected to extended reinforcement learning focused on difficult, multi-hour tasks. The result, according to xAI, is a system that works longer on hard problems, verifies its own outputs more carefully, manages extended context more effectively, and integrates natively with agent frameworks such as Grok Build.

What stands out immediately is the pricing decision. Input tokens remain at $2 per million and output tokens at $6 per million—the same rates as the preceding version. In an industry where each successive frontier model often arrives with higher costs, this continuity is notable. A faster variant, running at double the speed and double the price, is available in select environments including Cursor and Grok Build, but the standard model keeps the previous economics intact. Cached input tokens are further discounted, reinforcing the model’s appeal for high-volume or iterative workloads.

Technically, Grok 4.7 expands the parameter count substantially relative to its predecessor, reaching approximately 2.1 trillion parameters in some reports, an increase of roughly 40 percent from Grok 4.6’s 1.5 trillion. Training incorporated a longer reinforcement-learning phase weighted toward problems that require sustained execution over many hours. The model supports a 500,000-token context window, accepts both text and image inputs, and produces text outputs. Reasoning effort is configurable across low, medium, high (the default), and xhigh settings, giving users control over the depth-versus-latency tradeoff. Native tool use includes function calling, web search, X search, and code execution. Knowledge is current through approximately May 2026.

xAI highlights particular strength in self-verification and long-horizon task management. The model is said to pause more reliably to check intermediate results before proceeding, reducing the accumulation of errors in extended agent loops. Context compaction techniques further support prolonged interactions, while encrypted reasoning content is returned by default on the Responses API to preserve continuity across multi-turn sessions without additional configuration. These design choices align with the growing demand for models that can operate as persistent collaborators rather than single-shot responders.

The most significant advances can be summarized in the following numbered list of core improvements and results:

1. Larger base model with extended reinforcement learning focused on multi-hour tasks, enabling better persistence on difficult problems.  

2. Enhanced self-verification of outputs, allowing the model to check its own work more carefully before delivering final results.  

3. Improved long-context handling and context compaction for sustained agentic workflows.  

4. Native integration with Grok Build for agentic coding and knowledge-work tasks.  

5. Terminal-Bench performance nearly doubled to 38.0 percent compared with the previous version.  

6. Harvey legal agent benchmark score of 19.6 percent.  

7. CursorBench 4.0 score rising to 46.3 percent from 40.4 percent.  

8. EEBench electrical-engineering result of 64.0 percent and DeepSWE software-engineering score of 71.0 percent under high effort.  

9. AA Briefcase multi-hour office-work score advancing to 1,657.  

10. Unchanged pricing of $2 per million input tokens and $6 per million output tokens, with optional faster variants at double the rate.

Benchmark results released by the company show clear internal progress across these dimensions. Independent assessments, such as those from Artificial Analysis, place Grok 4.7 around 46 on their Intelligence Index—competitive but not leading the absolute frontier—while noting stronger relative standing on price-performance and coding-agent indices when paired with the Grok Build harness.

Comparisons with contemporaneous models reveal a familiar pattern. On several internal and third-party tests Grok 4.7 outperforms earlier OpenAI offerings such as GPT-5.6 Sol variants and remains cost-competitive against higher-priced systems like Claude Fable 5.1. It trails the absolute leaders on pure intelligence aggregates and certain pure agentic coding suites, yet its token economics—often described as half the price or better relative to comparable frontier models—make it attractive for sustained, high-volume use. One independent coding evaluation of Next.js tasks ranked it close behind top competitors while underscoring the substantial cost advantage. Early user reports from developers at Tesla and elsewhere describe productive overnight agent runs and efficient daily workflows when the model is paired with optimized harnesses that improve presentation, self-validation, and long-horizon monitoring.

Availability is immediate across multiple channels. The model is accessible through the xAI API under the identifier grok-4.7, inside the Cursor editor, via Grok Build, and through various third-party gateways and cloud platforms. Regional endpoints that keep inference within the United States are offered at a modest premium. Safety improvements are also claimed, with stronger performance on benchmarks designed to test resistance to malicious biology and cybersecurity requests.

The broader context of the release includes a rapid cadence of model updates throughout 2026. Earlier versions such as Grok 4.5 and 4.6 had already pushed coding and engineering capabilities forward, often incorporating specialized data from development environments and real-world engineering domains including SpaceX systems. Grok 4.7 continues that trajectory by emphasizing reliability over extended time horizons rather than pure peak intelligence on short benchmarks. For organizations running large numbers of coding agents, document-generation pipelines, or research workflows that span hours, the combination of improved persistence, self-checking, and stable pricing may prove more consequential than marginal gains on leaderboard scores.

Market reaction has been measured but positive among cost-conscious developers. Comments from practitioners highlight the model’s utility as a “daily workhorse” when integrated with agent teams that can parallelize planning, coding, and review. Demonstrations of complex outputs—such as interactive 3D visualizations generated from simple prompts—have circulated as illustrations of practical capability. At the same time, observers note that absolute state-of-the-art performance on the most demanding agentic and knowledge-work suites still belongs to a small set of higher-priced models. Grok 4.7’s contribution is therefore best understood as expanding the accessible frontier: bringing longer-horizon reliability and competitive coding strength to a price point that supports broader experimentation and production deployment.

Conclusion

Grok 4.7 represents a pragmatic advance in the ongoing race for capable, cost-effective AI systems. By delivering measurable gains in self-verification, long-horizon performance, and agentic reliability while holding pricing steady, xAI has strengthened its position for users who prioritize sustained productivity over maximum benchmark scores. The numbered improvements listed above illustrate a coherent focus on real-world utility rather than isolated peak results. As the model sees wider adoption across coding platforms, enterprise workflows, and agent frameworks, its true value will be measured by the volume of reliable work it enables at predictable cost. In a market increasingly defined by the balance between capability and economics, Grok 4.7 offers a clear, usable option that expands access to advanced AI assistance without forcing users to accept higher prices for incremental progress.