OpenAI has introduced GPT-5.6, but this release is not just another model replacing the previous one.
There are three new models: Sol, Terra and Luna. ChatGPT also has new reasoning controls, a separate Work mode, automatic model switching and different model availability depending on where and how you use the product.
This can be confusing at first.
Which model are you actually talking to? Why can you select Sol in a normal chat, but not Terra or Luna? What is the difference between ChatGPT and ChatGPT Work? And why does the model picker show reasoning levels instead of model names?
There is no single answer because ChatGPT is no longer a single chat interface running a single model. It is becoming a collection of interfaces, models and agents designed for different kinds of work.
Let’s see how the pieces fit together.
GPT-5.6 in a nutshell
GPT-5.6 is a family of three models:
- Sol is the flagship model for difficult reasoning and complex professional work.
- Terra balances capability, speed and cost.
- Luna is the fastest and least expensive model in the family.
The names describe stable capability tiers. The number describes the model generation.
This means that Sol, Terra and Luna may continue as product categories even when OpenAI moves beyond version 5.6. Instead of inventing a completely new set of model names for every release, OpenAI can update each tier independently.
You can think of the family like this:
- use Sol when quality is the most important requirement
- use Terra for regular production workloads
- use Luna when latency and cost matter more
The distinction is easy enough in the API, where developers explicitly select a model.
Inside ChatGPT, things are more complicated.
Does GPT-5.6 replace GPT-5.5?
Not completely.
GPT-5.5 Instant is still the default model for fast, everyday conversations. GPT-5.6 Sol is used when ChatGPT needs more reasoning, or when the user manually selects a higher reasoning level.
In standard ChatGPT conversations, the options now work roughly like this:
- Instant uses GPT-5.5 Instant
- Medium uses GPT-5.6 Sol
- High uses GPT-5.6 Sol with more reasoning
- Extra High uses GPT-5.6 Sol with the highest standard reasoning effort
- Pro uses GPT-5.6 Sol Pro
So the model picker is no longer only a model picker.
It is also a compute picker.
You are selecting how much time and processing ChatGPT should spend on the problem. Medium, High and Extra High may use the same underlying Sol model, but they do not necessarily use the same amount of reasoning.
This matters because model capability and reasoning effort are different variables.
A capable model answering immediately can be less useful than the same model spending more time planning, checking its work and using tools. On the other hand, using the maximum reasoning setting for a simple question wastes time and limited usage.
There is no reason to use Extra High to rewrite a two-sentence email.
There may be a good reason to use it to inspect a complex architecture, investigate a production issue or compare several implementation strategies.
Automatic reasoning
Eligible paid plans can allow ChatGPT to switch automatically from Instant to Medium when a request appears to require more reasoning.
For example, consider these two prompts:
- Convert 20 degrees Celsius to Fahrenheit.
and:
- Review this distributed job-processing architecture and identify failure modes that could cause duplicate execution.
The first request does not need a frontier reasoning model. The second may benefit from one.
When automatic switching is enabled, ChatGPT can make that decision without requiring the user to change the model manually. The interface may begin in Instant mode and move to Medium for the more difficult request.
This sounds simple, but it represents an important product change.
Previously, users were expected to understand the model lineup and select the correct model themselves. Now ChatGPT is beginning to act as a router that decides which kind of intelligence a task needs.
The long-term goal is probably not to make users better at selecting models.
It is to make model selection unnecessary.
Where are Terra and Luna?
You cannot select Terra or Luna in a standard ChatGPT conversation.
In regular chat, GPT-5.6 reasoning is handled by Sol. Terra and Luna are available in other parts of the OpenAI product line, including ChatGPT Work, Codex and the OpenAI API.
This separation makes more sense when we consider the type of work each interface performs.
A normal conversation is interactive. The user asks something, receives an answer and continues the discussion. For paid users, OpenAI can route difficult questions to Sol while keeping GPT-5.5 Instant as the faster default.
An agentic task is different.
An agent may search through files, browse websites, execute code, call tools, revise its plan and work for an extended period. Cost and latency can accumulate across dozens of model calls.
In this environment, choosing between Sol, Terra and Luna becomes more important.
A team might use:
- Sol for difficult planning and final review
- Terra for most intermediate work
- Luna for high-volume extraction, classification or formatting
Using the largest model for every step would be similar to running every microservice on the most expensive machine available. It works, but it is rarely the most efficient architecture.
What is ChatGPT Work?
ChatGPT Work is OpenAI’s interface for longer-running, multi-step tasks.
Instead of treating the interaction as a series of isolated questions, Work starts with a goal. The system can then plan the work, use tools, adapt when new information appears and produce a finished result.
A traditional chat request may look like this:
Give me five competitors in this market.
A Work task may look like this:
Research this market, identify the main competitors, compare their positioning and pricing, and create a report with the findings.
The first request mainly asks for an answer.
The second asks for an outcome.
That difference is important. When users ask for an outcome, the model must do more than generate a plausible response. It has to manage a process.
A simplified agent loop could look like this:
- Understand the goal.
- Break it into smaller tasks.
- Select tools and sources.
- Perform the tasks.
- inspect the intermediate results.
- Correct mistakes or change the plan.
- Produce the final artifact.
This is why the complete GPT-5.6 family is available in Work. Different stages of the loop may have different requirements.
Sol can handle the most ambiguous or difficult parts. Terra can provide a better balance for routine professional work. Luna can process simpler steps quickly and at a lower cost.
ChatGPT is becoming an orchestration layer
It is tempting to judge the GPT-5.6 launch only by comparing benchmark scores.
Sol performs better than previous OpenAI models on several agentic browsing, computer-use, coding and professional-work evaluations. OpenAI also reports that the new family can often complete tasks with fewer output tokens or tool calls.
These improvements matter, but the larger change is architectural.
ChatGPT now has several layers:
- an interface for communicating with the user
- a router for choosing a model and reasoning level
- models with different capability and cost profiles
- tools for searching, coding and working with files
- agent loops for performing multi-step work
- memory and project context that persist between interactions
The model is only one part of the system.
This is similar to the way a production application is more than the source code of one service. The database, queues, observability stack, deployment environment and communication between components are also part of the product.
In the same way, evaluating ChatGPT by asking which model it uses may no longer tell us enough.
We also have to ask:
- Which interface is running the task?
- How much reasoning effort is enabled?
- Which tools can the model access?
- Can it delegate work?
- Does it retain project context?
- Can it execute and verify actions?
- What happens when it reaches a usage limit?
Two users may both say that they used ChatGPT, while using substantially different systems.
What improved with GPT-5.6?
OpenAI is positioning GPT-5.6 around complex professional work rather than simple chatbot responses.
The announced improvements cover several areas.
Long-running tasks
Sol is designed to maintain focus across longer workflows. This is valuable for tasks in which the model has to research, plan, implement and revise instead of generating a single response.
The main challenge with long-running agents is not starting the task.
It is preserving the original goal after many intermediate steps.
A model may produce a good plan, call the correct tools and still drift toward an incomplete or slightly different result. Better persistence and instruction tracking can reduce this problem.
Computer and tool use
GPT-5.6 improves on evaluations involving web navigation and computer interaction. OpenAI reports a score of 62.6 percent for Sol on OSWorld 2.0 and 90.4 percent on BrowseComp, increasing to 92.2 percent with the Ultra reasoning setting.
Benchmark results should not be confused with guaranteed production performance. Real environments contain authentication problems, unusual interfaces, incomplete data and ambiguous user instructions.
Still, stronger tool use is necessary if ChatGPT Work is expected to complete tasks rather than only explain how they could be completed.
Documents, presentations and spreadsheets
OpenAI also highlights improved artifact generation.
GPT-5.6 is designed to follow reference documents and templates more accurately, including their layout, typography, hierarchy and recurring design rules. It can generate editable presentations, structured documents and spreadsheets instead of returning only plain text.
This changes the expected output of an AI assistant.
A useful answer may no longer be a paragraph explaining how to build a financial model. It may be the financial model itself.
Of course, generated artifacts still need review. A spreadsheet can look polished while containing an incorrect formula. A presentation can have consistent spacing while making an unsupported claim.
Better formatting does not remove the need for verification.
It raises the quality of the draft from which verification begins.
Availability depends on the product and plan
GPT-5.6 is not exposed uniformly across ChatGPT.
In standard conversations:
- Plus users receive Medium and High reasoning with Sol
- Pro, Business and Enterprise users also receive Extra High
- Pro is available on the Pro, Business and Enterprise plans
- Free and Go users do not receive GPT-5.6 Sol in standard chat
In ChatGPT Work:
- paid Plus, Pro, Business and Enterprise users can select Sol, Terra or Luna
In Codex:
- Free and Go users receive Terra
- Plus and higher plans can select Sol, Terra or Luna
In the API:
- developers can use all three models
GPT-5.5 Instant remains the everyday default in ChatGPT, and fallback models may be used after limits are reached.
The result is a product matrix rather than a simple model release.
When testing ChatGPT, it is worth recording the product mode, selected reasoning level and plan. Otherwise, two tests may not be comparable even when both appear to use “ChatGPT.”
What does this mean for developers?
Developers now have more control, but they also have more decisions to make.
Selecting a model should depend on the task rather than a general belief that bigger is always better.
Use Sol for uncertain and high-value work
Sol is the reasonable choice when errors are expensive, the task is ambiguous or the agent must maintain coherence across many steps.
Examples include:
- architecture reviews
- complex debugging
- security analysis
- research synthesis
- planning a large migration
- final verification of agent output
Use Terra for regular production tasks
Terra is intended to balance performance, speed and cost.
It may be suitable for:
- standard coding tasks
- document processing
- research with a clear scope
- data analysis
- internal workflow automation
- most steps inside an agent pipeline
Use Luna for speed and volume
Luna is the fastest and lowest-cost member of the family.
Possible uses include:
- classification
- extraction
- basic transformations
- formatting
- routing
- generating intermediate summaries
- processing large numbers of simple requests
The best system may use more than one model.
For example, Luna can classify incoming requests, Terra can perform the main work and Sol can review only the difficult or high-risk results.
This adds engineering complexity, but it can make agentic systems faster and cheaper without applying the same capability level to every task.
What about Codex hardware and ChatGPT Live?
OpenAI has also announced dedicated hardware for controlling Codex workflows, while GPT-Live introduces a new architecture for continuous voice interaction.
Both announcements support the same broader direction: AI systems are moving beyond the traditional prompt-and-response window.
The Codex hardware deserves a separate look because it introduces physical controls for supervising multiple coding agents.
GPT-Live also deserves its own article because it separates the conversational voice layer from the models performing deeper search and reasoning. The voice model can continue listening and speaking while delegating more complex work to another model.
For now, the relevant point is that GPT-5.6 is not an isolated release. It is one part of a larger reorganisation of ChatGPT around specialised interfaces and coordinated agents.
The new challenge: understanding the system
The old question was:
- Which ChatGPT model should I use?
The new questions are:
- Should this be a conversation or an agent task?
- How much reasoning does it require?
- Which parts need frontier intelligence?
- Which parts can run on a faster and less expensive model?
- How will the output be checked?
Sol, Terra and Luna provide more choices, but the choice of model is not the most important decision.
The important decision is how the models, tools and interfaces are combined.
There is no silver bullet here. Selecting Sol with the maximum reasoning level will not automatically turn an unclear request into a reliable workflow. Agents still need defined goals, access controls, observability, evaluation and human review.
GPT-5.6 makes the individual components more capable.
The real challenge is learning how to operate the system they are becoming part of.
Building AI products with GPT-5.6?
Understanding the new ChatGPT model lineup is only the first step. Turning those capabilities into reliable production systems requires the right architecture, tooling and engineering practices.
Whether you’re building AI agents, internal copilots, RAG systems or custom LLM-powered applications, RisingStack can help you design, build and scale them.
Learn more about our AI development services and talk to our team about your next AI project.
