When to Use GPT-5.6 Sol Terra or Luna in Microsoft Foundry Projects

In this blog post When to Use GPT-5.6 Sol Terra or Luna in Microsoft Foundry Projects we will explain how to match each model to the right business workload without paying for more AI capability than you need.

The common mistake is simple. A business deploys the most powerful model for every request, then wonders why costs and response times increase as usage grows. The opposite mistake is choosing the cheapest model for everything and discovering that important answers are incomplete, inconsistent or simply wrong.

What GPT-5.6 Sol, Terra and Luna actually do

GPT-5.6 is a family of OpenAI models available through Microsoft Foundry, which is Microsoft’s platform for building, securing and managing business AI applications. The three models use similar underlying technology, but they are designed for different levels of complexity, speed and cost.

All three take information such as text, documents or images and predict a useful response. They can also use approved business tools, search connected information and complete multi-step tasks. The main difference is how much computing effort they apply to understanding and completing the request.

• GPT-5.6 Sol is the flagship model for complex reasoning, detailed analysis and longer tasks involving multiple steps.
• GPT-5.6 Terra balances capability, speed and cost for everyday business applications.
• GPT-5.6 Luna is the fastest and most economical option for simpler, high-volume work.

Microsoft Foundry provides the business controls around these models. It lets your team manage access, connect approved data, monitor usage and choose where processing occurs. For Australian organisations, the Asia-Pacific Data Zone can help keep processing within the APAC region, although that should not be confused with an Australia-only data residency commitment.

If you want a broader model selection framework first, read our guide on choosing the right GPT-5.6 model for a business workload. Here, we will focus on practical deployment decisions inside Microsoft Foundry.

Use Luna for fast, repeatable and high-volume work

Luna is usually the right starting point when a task is clearly defined and repeated hundreds or thousands of times. These workloads do not need deep strategic reasoning. They need consistent results delivered quickly and economically.

Good examples include:

• Classifying support tickets by department or urgency
• Extracting names, dates and reference numbers from documents
• Summarising short emails or meeting notes
• Drafting standard customer service replies
• Tagging content for search and reporting
• Checking whether required fields are missing from a form

Consider an operations team processing 8,000 supplier emails each month. Luna could identify the supplier, purchase order number and requested action before sending the information to the appropriate workflow.

Using Sol for this work would be like hiring a senior commercial lawyer to sort the office mail. The result might be excellent, but the extra capability provides little business value.

Luna is not the best choice when a request is ambiguous, carries significant financial risk or requires the model to compare conflicting evidence. In those cases, the apparent saving can disappear through rework and human corrections.

Use Terra as the default for everyday business AI

For many organisations, Terra should handle the majority of employee-facing AI requests. It provides stronger reasoning than Luna without applying the full cost and processing effort of Sol to every interaction.

Terra is well suited to:

• Internal assistants that answer questions from policies and procedures
• Drafting reports, proposals and customer communications
• Summarising longer documents with business context
• Preparing management briefings from several information sources
• Assisting service desk and operations teams with recommended actions
• Running structured workflows that call approved business systems

A 200-person professional services firm might use Terra for an internal assistant connected to SharePoint. Employees could ask about leave policies, project templates or client onboarding procedures without searching through multiple folders.

The business outcome is not simply faster answers. It is fewer interruptions for HR, finance and IT teams, along with less time spent looking for information.

Terra can also be a sensible first model for a pilot. Once you collect real usage data, simpler requests can move to Luna while difficult requests are escalated to Sol.

Use Sol when the cost of a weak answer is high

Sol is designed for work where deeper analysis, planning and verification matter more than receiving an instant answer. It is particularly useful when the model must handle uncertainty, use several tools or maintain context across a longer workflow.

Suitable use cases include:

• Comparing complex contracts, policies or technical proposals
• Investigating a cybersecurity incident across multiple evidence sources
• Planning a cloud migration with dependencies and risk controls
• Reviewing large software systems or generating complex code
• Coordinating AI agents that complete separate parts of a larger task
• Preparing detailed decision papers for senior management

Imagine a CIO asking an AI assistant to compare three cloud proposals, identify hidden dependencies, assess security gaps and prepare questions for each supplier. This is not basic summarisation. The model must connect evidence, recognise uncertainty and produce a defensible recommendation.

Sol is the stronger candidate because a shallow answer could influence a costly business decision. Human review should still remain mandatory for legal, security, employment and major financial decisions.

For development teams, Sol may also support complex engineering work. However, specialist coding models can still make sense for particular workflows, as discussed in our article on what IT teams should do with GPT-5.3 Codex.

The best design usually uses all three models

You do not need to select one GPT-5.6 model for the entire organisation. Microsoft Foundry allows applications to use separate model deployments, so each request can be sent to the model that best fits its risk and complexity.

For example, a service management assistant could use:

• Luna to classify incoming requests and extract key details
• Terra to draft responses using approved knowledge articles
• Sol to investigate unusual incidents involving several systems

This approach controls costs without lowering quality where it matters. It also avoids forcing employees to understand model names before asking for help.

A simple application-level routing pattern could look like this:

def choose_model(task):
 if task.risk == "high" or task.complexity == "complex":
 return "gpt-56-sol-production"

 if task.volume == "high" and task.complexity == "simple":
 return "gpt-56-luna-production"

 return "gpt-56-terra-production"

response = client.responses.create(
 model=choose_model(task),
 input=task.instructions
)

The names in this example represent deployments created in your Azure environment. In production, routing should also consider response quality, permitted data, budget limits and whether human approval is required.

Test cost per successful task rather than token price

AI services are generally charged according to the amount of information processed and generated. These units are called tokens, which are small pieces of words and punctuation.

A cheaper model is not cheaper if employees must repeat requests or correct poor output. Likewise, the most capable model is wasteful if a simpler option completes the task correctly.

Before deployment, test 30 to 50 representative examples and measure:

1. How often the model produces an acceptable result
2. How long users wait for a response
3. The average cost of each completed task
4. How often the request needs human correction
5. Whether sensitive information is handled appropriately

Do not rely on public leaderboards alone. Your policies, documents, customer language and business processes are the real test set.

Governance matters as much as model selection

Microsoft Foundry can manage identity, deployment locations, usage monitoring and access to connected tools. These controls help prevent an AI assistant from seeing information or taking actions beyond the user’s authority.

Australian organisations should also consider Privacy Act obligations, contractual data requirements and internal information classifications. The Essential Eight, the Australian Government’s baseline cybersecurity framework, does not certify an AI model, but its controls remain relevant to the devices, identities and applications surrounding the AI service.

CloudProInc approaches these projects as both an AI and security exercise. Our experience across Microsoft Azure, Microsoft 365, Intune, Defender and Wiz helps clients address the full environment rather than treating the model as an isolated tool.

For a wider view of providers and governance, see our guide to selecting OpenAI, Claude and Meta models in Microsoft Foundry.

Start with Terra and earn the right to move

If you have no production evidence yet, Terra is often the practical starting point. Move repetitive, low-risk requests to Luna after testing, and move genuinely difficult or high-impact work to Sol.

The goal is not to deploy the smartest model everywhere. It is to achieve the right business result at a predictable cost, with security controls that stand up when the pilot becomes a company-wide service.

If you are unsure whether Sol, Terra or Luna fits your Microsoft Foundry project, CloudProInc can help assess the workload, run a controlled comparison and estimate production costs. We are a Melbourne-based Microsoft Partner and Wiz Security Integrator with more than 20 years of hands-on enterprise IT experience, and we are happy to take an initial look without turning it into a sales exercise.