Grok 4.7 Ties Xiaomi MiMo-V2.6-Pro on Intelligence Index at Five Times Token Price

Grok 4.7 is SpaceXAI's most capable model for coding and knowledge work, released on September 21, 2026, using a new larger base model and longer reinforcement learning run.

The release of Grok 4.7 by SpaceXAI on September 21, 2026, marks a significant step in the development of frontier models tailored for coding and knowledge work. The model builds upon previous iterations by employing a larger base model and an extended period of reinforcement learning. This approach allows the model to handle difficult tasks for longer durations and to verify its outputs with greater care. SpaceXAI has also incorporated improved safeguards in this version. The timing of the release coincides closely with Xiaomi's introduction of its own advanced model, creating a direct point of comparison in the current landscape of AI development. Industry observers note that such simultaneous launches reflect the accelerating pace of innovation where multiple organizations push boundaries in parallel.

Xiaomi launched MiMo-V2.6-Pro on September 21-22, 2026, as a 1.02-trillion-parameter omnimodal mixture-of-experts model. The model activates 42 billion parameters during inference and supports a context window of 1 million tokens. It is distributed under the MIT license with weights hosted on Hugging Face, making it accessible for a wide range of users and developers. This open-weight approach contrasts with the closed nature of Grok 4.7, which is served through SpaceXAI's API. The omnimodal capabilities allow the model to process multiple types of input, expanding its potential applications beyond text-based tasks. Community developers have already begun experimenting with fine-tuning the released weights for specialized domains.

Both models have been evaluated on the Artificial Analysis Intelligence Index, where they achieve nearly identical scores. Grok 4.7 (xhigh) scores 46, while MiMo-V2.6-Pro scores 46.32, but the comparison places both at 46. This benchmark provides a composite measure of intelligence and performance across various tasks. The close scores suggest that open-weight models are rapidly closing the gap with proprietary frontier models in terms of raw capability. Analysts attribute this convergence to improvements in training data quality and architectural efficiency rather than sheer scale alone.

How do the technical specifications of Grok 4.7 and MiMo-V2.6-Pro compare?

The technical architecture of MiMo-V2.6-Pro features a mixture-of-experts design with a total of 1.02 trillion parameters, of which 42 billion are active at any given time. This sparse activation allows for efficient computation while maintaining high performance. The 1 million token context window enables the model to handle very long inputs, which is useful for tasks involving large documents or extensive codebases. In contrast, details on the exact parameter count for Grok 4.7 have not been disclosed by SpaceXAI. The model is described as using a new larger base model, but specific figures remain proprietary. Experts suggest that undisclosed details may include custom optimizations for coding workflows.

Performance on specific benchmarks further illustrates the capabilities. Grok 4.7 achieves a score of 46.3 percent on CursorBench 4.0, a metric focused on coding proficiency. This result highlights its strength in coding tasks, aligning with SpaceXAI's description of it as the most capable model for such work. MiMo-V2.6-Pro, being omnimodal, likely excels in a broader range of modalities, though specific benchmark scores beyond the intelligence index are highlighted in Xiaomi's announcements. The combination of high context length and active parameter efficiency positions MiMo-V2.6-Pro as a versatile option for research and production environments.

Key specifications and pricing for Grok 4.7 and MiMo-V2.6-Pro

Model	Total Parameters	Active Parameters	Context Window	Input Price per Million Tokens	Output Price per Million Tokens	Intelligence Index Score	License
Grok 4.7	Undisclosed	Undisclosed	Undisclosed	2	6	46	Proprietary
MiMo-V2.6-Pro	1.02 trillion	42 billion	1 million tokens	0.435	0.87	46.32	MIT

The pricing difference is substantial. Grok 4.7 is offered at $2 per million input tokens and $6 per million output tokens. This pricing structure is the same as the previous Grok 4.6 model. In comparison, MiMo-V2.6-Pro is available through the Xiaomi API at $0.435 per million input tokens and $0.87 per million output tokens. This represents approximately four to seven times lower cost depending on the token type, providing a significant economic advantage for users of the open-weight model. Cost-sensitive organizations may therefore favor the Xiaomi offering for high-volume inference workloads.

What are the market implications of these competing releases?

The simultaneous releases underscore the competitive nature of the frontier AI market. Open-weight models like MiMo-V2.6-Pro are challenging the dominance of closed models by offering comparable performance at lower prices. This shift could accelerate adoption in industries where cost is a primary factor. SpaceXAI maintains its position with proprietary features and integration, but the pressure from open alternatives is evident. Market analysts predict that the availability of high-performing open models will influence procurement decisions across sectors including software development and data analysis.

Stakeholders including researchers, businesses, and individual developers must weigh the trade-offs between performance, cost, accessibility, and control. The availability of weights for MiMo-V2.6-Pro enables customization and experimentation that API-only models do not permit. However, Grok 4.7 may offer advantages in areas such as safety calibration and task-specific optimization for coding. Hybrid strategies that combine both models for different use cases are likely to emerge as organizations seek to balance these factors.

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

Expert reactions have focused on the narrowing performance gap between open and closed models. The fact that an open-weight model can match a frontier model on a standardized intelligence index suggests that innovation is occurring across the spectrum. This may lead to more hybrid approaches where organizations use a combination of open and proprietary tools. Discussions in technical forums emphasize the potential for community-driven improvements to open models that could further close any remaining gaps in specialized tasks.

What developments can be expected in the coming months for these models?

1. Further optimization of mixture-of-experts architectures to improve efficiency across both open and closed systems.
2. Expansion of context windows beyond 1 million tokens in subsequent releases from multiple providers.
3. Increased focus on multimodal capabilities across both open and closed models to handle diverse data types.
4. Potential updates to pricing structures as competition intensifies and new entrants join the market.
5. More detailed benchmark disclosures from model providers to allow finer-grained comparisons.

Looking ahead, the AI community will monitor how these models evolve and how the market responds to the pricing disparity. The open release of MiMo-V2.6-Pro could inspire similar moves from other companies, potentially democratizing access to high-performance AI. SpaceXAI may respond with enhancements to Grok 4.7 or new models that justify the premium pricing through superior performance in niche areas. Continued investment in reinforcement learning techniques is expected to drive incremental gains in both offerings.

The overall trend indicates a maturing field where performance metrics are converging, but business models and access methods continue to differentiate offerings. Users benefit from increased choice, while providers must innovate continuously to maintain relevance. This dynamic is likely to persist as more players enter the frontier model space. Regulatory scrutiny may also increase as open models proliferate, prompting discussions on safety standards and deployment guidelines.

Additional context from industry reports highlights that the 1.02 trillion parameter scale achieved by MiMo-V2.6-Pro represents a milestone for open-weight systems. Such scale was previously associated primarily with closed models. The ability to serve this model at low token prices through the Xiaomi API demonstrates advances in inference optimization. These factors collectively point to a future where high capability is no longer exclusively tied to high cost.

Developers working with Grok 4.7 benefit from its described longer task handling and self-verification features, which are particularly valuable in professional coding environments. Meanwhile, the MIT license for MiMo-V2.6-Pro supports commercial use and modification without restrictive terms. Both paths contribute to a diverse ecosystem that caters to varying organizational needs and risk tolerances.