> Blog >

Grok 4.7 and GLM 5.3 Join Cortex AI: Expanding Frontier Model Choice on Snowflake

Grok 4.7 and GLM 5.3 Join Cortex AI: Expanding Frontier Model Choice on Snowflake

Fred
November 20, 2026

Enterprise AI strategy has moved decisively beyond single-model dependence. The highest-performing organizations treat models as interchangeable capabilities that can be selected—or automatically routed—according to task requirements, cost targets, latency budgets, and risk posture. On October 16, 2026, Snowflake expanded that choice inside Cortex AI by making two additional frontier models available in Public Preview: grok-4.7 from xAI and glm-5.3 from Z.ai.

Both models are accessible across AI Functions, Cortex Inference, Cortex Agents, Snowflake CoWork, and Snowflake CoCo (with availability nuances by product). They support cross-region inference when CORTEX_ENABLED_CROSS_REGION is set appropriately. The additions strengthen Snowflake’s multi-model posture and give teams more options for coding, knowledge work, and long-running agentic workloads—while keeping data inside the governed Snowflake perimeter.

This post examines the strengths each model brings, availability details, expected performance characteristics, ideal workloads, the governance advantages of running models inside Cortex, how they complement the existing catalog, competitive context, implications for model-routing strategies, and what the expansion signals for enterprise AI flexibility in late 2026.

Snowflake’s Multi-Model Strategy

Snowflake has consistently positioned Cortex AI as a multi-model platform rather than a single-model service. The goal is to let customers choose (or let the platform help choose) the right model for each job while applying consistent security, governance, observability, and cost controls. Adding Grok 4.7 and GLM 5.3 continues that trajectory: more frontier and near-frontier options, available next to governed enterprise data, without requiring separate model hosting or data movement.

Model Highlights

Grok 4.7 (xAI)

  • Positioned as a highly capable model for coding and knowledge work.
  • Designed to sustain longer, more difficult tasks and to check its own work more carefully.
  • Competitive price-performance profile relative to other frontier models in its class.
  • Strong results reported on agentic coding and longer-horizon knowledge-work benchmarks.
  • Configurable reasoning effort levels in its native serving environments.
  • Context window on the order of hundreds of thousands of tokens (native specifications cite 500K).

GLM 5.3 (Z.ai)

  • Flagship open-weight model focused on complex software engineering and agentic tasks.
  • Mixture-of-Experts architecture with a very large total parameter count and substantial active parameters per token.
  • Large context window (native specifications cite 1M tokens) and high maximum output length.
  • Notable strength in coding, long-horizon multi-step engineering tasks, and emerging cyber/security-related capabilities.
  • Selectable reasoning effort levels that trade latency and token consumption against task performance.
  • Attractive option for teams that value strong open-weight alternatives alongside proprietary frontier models.

Both models enter Cortex in Public Preview; production suitability and regional nuances should be validated against current documentation and account configuration.

Availability and Integration Points

According to the official release notes, grok-4.7 and glm-5.3 are available in Public Preview for:

  • Cortex AI Functions
  • Cortex Inference
  • Cortex Agents
  • Snowflake CoWork
  • Snowflake CoCo

Cross-region inference is supported when the account’s cross-region setting permits it (ANY_REGION, AWS_GLOBAL, or AWS_US as documented). Teams should confirm exact regional availability and any product-specific limitations (for example, evaluation or certain agent surfaces) in the latest model-availability tables.

Because the models are accessed through Cortex, they inherit the same authentication, role-based access, network policies, and audit logging as other Cortex model invocations. Data used in prompts or retrieved via tools remains subject to Snowflake’s existing governance controls.

Expected Performance Characteristics and Ideal Workloads

While independent, apples-to-apples benchmarks inside every customer environment will vary, the public positioning of the two models suggests complementary strengths:

Grok 4.7 is well suited to:

  • Agentic coding workflows and longer software-engineering tasks
  • Knowledge work that benefits from sustained reasoning and self-checking
  • Interactive and multi-step analytic or research agents
  • Scenarios where strong price-performance at the frontier matters

GLM 5.3 is well suited to:

  • Complex, repository-scale or multi-file coding and refactoring
  • Long-horizon agentic engineering tasks that involve many tool calls
  • Workloads that benefit from a very large context window
  • Teams seeking a high-capability open-weight option for coding and agentic use cases
  • Exploratory or cost-sensitive workloads that can take advantage of open-model economics

In practice, many organizations will evaluate both models on representative internal tasks (code generation and repair, multi-step data analysis, document synthesis, tool-using agents) and then encode preferences into routing logic or agent configurations.

Governance Advantages of Running Inside Cortex

Running frontier models inside Cortex rather than through external endpoints delivers several structural advantages:

  • Data proximity — Prompts and retrieved context can stay inside the Snowflake security perimeter.
  • Unified access control — The same roles, row-access policies, and masking policies that protect tables also constrain what agents and AI Functions can see.
  • Consistent observability — Usage, cost, and (where instrumented) traces feed into the same monitoring surfaces used for other Cortex activity.
  • Centralized policy — Model enablement, cross-region settings, and guardrails can be managed at the account or organizational level.
  • Reduced integration surface — Applications and agents call a single Cortex interface rather than maintaining separate credentials and SDKs for every model provider.

These advantages become more valuable as the number of models in active use grows. Multi-model strategies only scale if governance and operations scale with them.

How the New Models Complement the Existing Catalog

Cortex already hosts a range of models from multiple providers, spanning different capability tiers, cost points, and specializations. Grok 4.7 adds another strong frontier option optimized for coding and sustained agentic work. GLM 5.3 adds a high-capability open-weight alternative with a large context window and demonstrated strength on engineering and agentic benchmarks.

Together they expand the “efficient frontier” available to routing layers and to human designers of agent systems. A typical portfolio might now include:

  • Highest-capability proprietary models for the most demanding reasoning
  • Strong coding- and agent-specialized models (including Grok 4.7)
  • High-value open-weight models (including GLM 5.3) for cost-sensitive or high-volume workloads
  • Smaller/faster models for classification, extraction, and simple generation

Dynamic or rules-based routing can then direct each request to an appropriate tier.

Competitive Context and Model-Routing Implications

The broader market continues to see rapid releases from both closed and open model labs. Enterprises that lock themselves to a single provider risk both capability gaps and pricing power disadvantages. Platforms that offer curated, governed access to multiple leading models—and that invest in routing, evaluation, and cost controls—reduce that risk.

For AI platform teams, the practical implications are clear:

  • Maintain an internal evaluation harness that regularly scores candidate models on representative tasks.
  • Encode routing policies that consider quality, cost, latency, and data-sensitivity requirements.
  • Treat model choice as a configurable property of agents and applications rather than a hard-coded dependency.
  • Monitor actual usage and cost by model so that routing rules can be refined with real data.

Snowflake’s continued expansion of the Cortex model catalog, including these two additions, supports exactly that operating model.

What This Signals for Enterprise AI Flexibility in Late 2026

By late 2026, model diversity inside governed platforms is becoming a baseline expectation rather than a differentiator. The addition of Grok 4.7 and GLM 5.3 reinforces three broader trends:

  1. Frontier capability is no longer monolithic — Different models lead on different axes (coding, long-horizon agency, context length, cost efficiency).
  2. Open-weight models remain competitive — High-capability open models continue to close gaps and expand the set of viable production options.
  3. Governance and routing matter as much as raw model quality — The winning architecture is often “best model for the task, under consistent controls,” not “single best model for everything.”

Organizations that build flexible, observable, multi-model stacks will adapt faster as the model landscape continues to shift.

Actionable Insights for AI Platform Teams

  • Add grok-4.7 and glm-5.3 to your internal model evaluation suite and score them on coding, multi-step agent, and knowledge-work tasks that matter to your business.
  • Confirm Public Preview status, regional availability, and any product-specific limitations in your account before promoting the models to production traffic.
  • Update agent configurations and routing rules to include the new models where they demonstrate advantage; start with shadow or canary traffic.
  • Leverage Cortex’s unified governance so that expanding model choice does not expand the compliance surface.
  • Track cost and quality metrics by model and by workload type; use the data to refine routing and to set expectations with application teams.
  • Document clear guidance for developers: when to prefer each model class, how to request new models, and how to handle model deprecation or preview-to-GA transitions.
  • Pair model diversity with strong evaluation and observability practices—more models only help if you can measure which ones actually perform better on your data and tasks.
  • Consider open-weight options such as GLM 5.3 for high-volume or cost-sensitive paths while reserving the highest-capability proprietary models for the most demanding cases.
  • Align with security and compliance stakeholders early so that cross-region inference settings and data-handling policies remain well understood.
  • Treat the Cortex model catalog as a living portfolio: regularly retire under-used models and onboard new ones through a lightweight but consistent process.

Grok 4.7 and GLM 5.3 joining Cortex AI in Public Preview expand the practical choices available to teams building on the AI Data Cloud. One brings frontier strength in coding and sustained knowledge work; the other brings high-capability open-weight performance with a large context window and strong agentic engineering characteristics. Both arrive inside the same governed, observable interface that already hosts Snowflake’s multi-model strategy. For enterprises serious about flexibility, cost discipline, and data control in late 2026, that combination—more strong models, one consistent control plane—is exactly what a mature AI platform should deliver.