GPT-5.6 Comes to ChatGPT | Sol, Terra, and Luna Explained, Plus Plan Availability and More

Now that “GPT-5.6” has started appearing in ChatGPT, you may be wondering how it differs from GPT-5.5—or whether you should choose Medium or High.

In short, GPT-5.6 is not primarily designed to answer everyday questions faster. It is built to reason more deeply through complex tasks such as programming, research, data analysis, scientific work, and computer use.

That does not mean every ChatGPT response now comes from GPT-5.6. GPT-5.5 Instant still handles everyday questions, while GPT-5.6 Sol powers thinking settings such as Medium, High, and Extra High.

Drawing on OpenAI’s official announcement and Help Center as of July 10, 2026, this article explains GPT-5.6’s features, the differences among its three models, plan availability, how to choose a thinking level, and the main points to keep in mind.

What Is GPT-5.6?

GPT-5.6 is a new model family that OpenAI announced as a limited preview on June 26, 2026. It has since been rolling out gradually to eligible ChatGPT plans.

According to official information, GPT-5.6 is designed for complex tasks requiring multiple steps, such as software development, specialized knowledge work, scientific research, cybersecurity, computer operations, and design.

It is designed to do more than answer questions: it can break problems into steps, gather the information it needs, use tools, and verify results as it works.

This information comes from OpenAI’s June 26, 2026 announcement and its official Help Center, updated in July 2026.

Its main strength is the ability to reason more deeply through complex instructions and lengthy tasks. That makes it useful for modifying software, comparing multiple documents, organizing long-form content, and performing specialized data analysis.

The tradeoff is that higher thinking levels take longer to respond and use up your allowance faster. Simple questions rarely need the highest setting.

In the interface, you choose a thinking level such as “Medium” or “High” instead of selecting “GPT-5.6” by name. In practice, the more useful question is how deeply you want ChatGPT to reason, not which model name appears behind the setting.

A practical approach is to use “Instant” or automatic switching for everyday questions, then move to “Medium” or above only for demanding development or research tasks.

Main Application Areas for GPT-5.6

• Software development and programming
• Specialized document creation and knowledge work
• Research and analysis using multiple sources
• Scientific research and data analysis
• Computer use and automation
• Support for design and production processes
• Defensive cybersecurity
• Complex, long-duration workflows

Differences Between GPT-5.6 Sol, Terra, and Luna

The GPT-5.6 family has three model tiers: Sol, Terra, and Luna.

According to OpenAI’s announcement, “5.6” identifies the model generation, while Sol, Terra, and Luna form a hierarchy based on capability, speed, and cost.

The top-tier Sol model is designed for deep analysis of complex problems. Terra prioritizes a balance of performance, speed, and cost, while Luna is positioned as the fastest and most affordable model.

This makes it easier to match capability and cost to the job. Rather than using the top-tier model for everything, you can choose Terra for routine work and Luna for lightweight processing.

The limitation is that standard ChatGPT does not let you freely choose among all three. As of July 10, 2026, regular ChatGPT conversations primarily use Sol, while Terra and Luna are mainly available through Work, Codex, and the OpenAI API.

Sol is not automatically the best choice for every task. For a quick sentence edit or a simple question, GPT-5.5 Instant or a lightweight model such as Luna may be more efficient.

Sol is well suited to investigating site-wide issues or revising multiple files. For a basic code explanation, a short proofreading pass, or another routine task, however, a higher thinking setting may make little noticeable difference.

The key is to weigh task difficulty, response time, and usage limits, then avoid choosing more power than the task requires.

Positioning of the Three Models

• Sol is the top-tier model in the GPT-5.6 series
• Sol is designed for complex development, investigation, research, and long-duration tasks
• Terra prioritizes a balance of capability, speed, and cost
• Terra is designed for everyday development and business tasks
• Luna is the fastest model in the GPT-5.6 series
• Luna is designed for low-cost, high-volume, light-duty processing
• Sol is primarily used for standard ChatGPT conversations
• Terra and Luna are available on Work, Codex, and via the API

GPT-5.5 Instant Will Continue to Be Used in ChatGPT

The release of GPT-5.6 does not mean it has replaced every model used in ChatGPT.

The official OpenAI Help Center explains that GPT-5.5 Instant will continue to be used as the default model for fast, everyday responses. GPT-5.6 Sol handles settings that require deeper thinking, such as Medium, High, and Extra High.

The idea is simple: GPT-5.5 Instant handles everyday conversations quickly, while GPT-5.6 Sol takes on complex tasks that call for deeper reasoning.

This approach keeps simple answers fast while reserving the more capable model for difficult problems, providing a better balance of speed and accuracy than routing every request through a resource-intensive model.

One source of confusion is that the active model can change within the same conversation. You generally cannot tell from the answer alone whether GPT-5.5 or GPT-5.6 produced it.

So it would be incorrect to assume that GPT-5.6 is always used by default simply because it has been released. GPT-5.5 Instant may still provide standard, fast responses.

When automatic switching is enabled, ChatGPT can move from Instant to Medium for more complex requests. You can turn this behavior on or off under “Configure” in the model selector.

Enable automatic switching if you do not want to choose manually. If you prefer to decide case by case, turn it off and select “Medium” or “High” only when needed.

The Roles of GPT-5.5 and GPT-5.6

• GPT-5.5 Instant quickly processes everyday questions
• GPT-5.6 Sol handles complex reasoning and lengthy tasks
• With automatic switching, the system transitions from Instant to Medium
• Automatic switching can be toggled via “Configure”
• GPT-5.6 does not completely replace GPT-5.5 Instant
• For simple questions, Instant is often sufficient

Differences Between Medium, High, Extra High, and Pro

GPT-5.6 lets you adjust not only the model but also how much compute it devotes to reasoning.

According to the official OpenAI Help Center, “Medium” is for standard reasoning, “High” is for longer reasoning, and “Extra High” is the highest reasoning setting available with GPT-5.6 Sol. “Pro” uses GPT-5.6 Sol Pro and is intended for particularly difficult problems or tasks requiring extended processing time.

Because Medium already uses GPT-5.6 Sol, most development work and in-depth research do not need to start at High or Pro.

Being able to match the thinking level to the task helps conserve your allowance and reduce wait times. A sensible approach is to switch to High only when Medium falls short.

More thinking does not guarantee a correct answer. If the instructions are vague or key information is missing, even Extra High or Pro may fail to produce the result you expect.

Thinking level does not control response length. Selecting “Extra High” will not necessarily produce more text; it gives the model more time and compute to analyze the problem internally.

In practice, clear instructions at Medium are often faster and more reliable than sending a simple correction to Extra High. Model capability matters, but so does describing the task precisely.

Start with Medium. If the result is not good enough, explain what is missing in the same conversation and then switch to High.

How to Choose a Thinking Level

• “Instant” is for short questions and everyday conversation
• “Medium” is for general research and development tasks
• “High” is for complex tasks involving multiple conditions
• Extra High is for design decisions and large-scale analysis
• Pro is for particularly difficult problems or tasks requiring a long time
• If you’re unsure, start with Medium
• Switch to High or higher only if the results are insufficient
• Use Instant when response speed is a priority

What Features Are Available with ChatGPT Plus?

As of July 10, 2026, the official OpenAI Help Center states that ChatGPT Plus users can access the Medium and High levels of GPT-5.6 Sol.

However, Extra High and Pro are not included in Plus. Extra High and Pro are primarily available with ChatGPT Pro, Business, and Enterprise.

GPT-5.6 Sol is not available in standard ChatGPT conversations on the Free and Go plans. However, the official documentation states that Free and Go users can access GPT-5.6 Terra in Codex.

Plus still provides access to GPT-5.6 Sol’s core capabilities through Medium and High. For personal development projects, document creation, research, and code review, that may be all you need.

What Plus does not include is access to the top-tier Extra High and Pro options. And because the rollout is phased, GPT-5.6 may not appear immediately even on an eligible account.

Even with a Plus subscription, GPT-5.6 Sol will not appear in the model selector until it reaches your account. In managed Business or Enterprise workspaces, administrators may also restrict model availability.

If you see “Medium” and “High,” you can use GPT-5.6 Sol with Plus. The interface may not prominently display the name “GPT-5.6,” but choosing either setting activates Sol.

Open the model selector and look for “Medium” and “High.” If they are missing, confirm that you are signed in to the right account and wait for the rollout to reach you.

Main Usage Scope by Plan

• Plus users can use Medium and High
• Plus users cannot use Extra High or Pro
• Pro users can use Medium, High, Extra High, and Pro
• Business plans can use Medium, High, Extra High, and Pro
• Enterprise plans can use Medium, High, Extra High, and Pro
• Free and Go plans cannot use GPT-5.6 Sol in standard conversations
• Free and Go plans can use Terra in Codex
• Even for eligible plans, availability may vary due to a phased rollout

Available Models Differ Between Work and Regular Chat

To understand where GPT-5.6 is available, you first need to distinguish regular chat from Work.

According to the official OpenAI Help Center, standard ChatGPT conversations primarily use Sol. In Work, users on Plus, Pro, Business, and Enterprise can choose among Sol, Terra, and Luna.

Standard ChatGPT is designed for conversational activities such as asking questions, getting advice, and drafting text. Work is an environment for multi-step tasks that use tools to produce a finished deliverable.

Work lets you choose the model that best fits the task: Sol for deep analysis, Terra for a balance of speed and capability, or Luna for lighter processing.

Because Chat and Work have different interfaces and workflows, the distinction may be confusing at first. Work tasks can also require more compute and credits than regular conversations.

Work limits and credit usage vary by plan, task, and workspace settings, so not every operation consumes the same amount of resources.

In general, regular chat is enough when you simply need an answer. Work is usually the better fit when the job involves files, multiple steps, or a concrete deliverable.

A useful rule of thumb is to use chat for questions and research, and Work for hands-on jobs such as revising a website or creating a document.

How to Choose Between Chat and Work

• Use regular chat for short questions or consultations
• Detailed research and writing can also be handled via regular chat
• Use Work for tasks involving multiple steps
• Use Work for tasks that involve files or tools
• Regular chat primarily uses Sol
• In Work, you can choose between Sol, Terra, and Luna
• Use Sol for tasks requiring high-level capabilities
• Use Terra for routine tasks
• Use Luna for light, high-speed processing

How to Use GPT-5.6 in Codex

GPT-5.6 is also available in Codex, OpenAI’s coding environment.

According to the official OpenAI Help Center, Sol, Terra, and Luna are available in Codex for Plus, Pro, Business, and Enterprise plans. Terra is available for Free and Go plans.

GPT-5.6 Sol is stronger at development tasks that require planning and iteration, including command-line work, changes across multiple files, and the investigation, repair, and testing of bugs.

OpenAI’s official announcement states that GPT-5.6 Sol achieved high scores on Terminal-Bench 2.1, a benchmark that evaluates command-line tasks. However, only a portion of the evaluation results has been disclosed at this time; the full results will be released when the model becomes generally available.

Its strengths go beyond generating code. It can read an existing project, investigate root causes, make coordinated fixes in several places, and verify the result.

Higher-capacity models generally require more compute, however, and Codex-generated changes are not always correct. Critical projects still require human review.

GPT-5.6 also has minimum version requirements in Codex. According to the official Help Center, Codex mode in the ChatGPT desktop app requires version 26.707.30751, while the Codex CLI requires version 0.144.0 or later.

You do not need Sol for every small visual adjustment or basic code explanation; Terra or a lower thinking setting may be enough. Sol is most valuable for unexplained bugs, large refactors, and architectural changes spanning multiple files.

For a safer workflow, begin with: “Investigate only; do not make changes yet. Identify what needs to change and assess the risks.” Review the findings before asking Codex to implement the fix.

Tasks Well-Suited for Codex

• Investigating bugs that span multiple files
• Identifying the causes of errors
• Implementing fixes and testing them
• Large-scale refactoring
• Development tasks using the command line
• Reviewing existing code
• Identifying security vulnerabilities
• Adding features based on specifications

GPT-5.6’s Scientific Research and Analytical Capabilities

Scientific research and complex data analysis are key focus areas for GPT-5.6.

According to research materials released by OpenAI around July 2, 2026, GPT-5.6 Sol achieved a pass rate of 28.7% on the “Thinking” setting and 31.5% in “Pro” mode on GeneBench-Pro, a benchmark designed to measure complex analytical capabilities in the life sciences.

This evaluation measures not just simple knowledge-based questions, but the ability to identify issues in the data, select appropriate analytical methods, and carry out multi-step processing.

This creates opportunities to accelerate parts of research and analysis that normally take experts considerable time. GPT-5.6 can help organize data, propose analytical methods, write code, and explain results.

Even in Pro mode, however, the pass rate is only 31.5%. The model is far from solving every problem correctly, and OpenAI acknowledges that current AI agents are not reliable enough to replace human experts.

In high-stakes fields such as healthcare, law, finance, and research, GPT-5.6’s output should never serve as the sole basis for a final decision. Expert review and verification against primary sources remain essential.

GPT-5.6’s value is not that it eliminates the need for experts. It can help organize the issues they need to consider and speed up the early stages of analysis.

Rather than asking only for a result, have the model report its assumptions, excluded data, chosen methods, alternative interpretations, and confidence level as well.

Items to Verify Through Expert Analysis

• Sources and data used
• Assumptions underlying the analysis
• Methods adopted
• Methods not adopted
• Missing or anomalous data
• Uncertainties affecting the conclusion
• Possibility of alternative interpretations
• Areas requiring human verification

Cybersecurity Capabilities and Safety Measures

OpenAI describes GPT-5.6 as a major step forward in cybersecurity capabilities.

According to the official announcement, GPT-5.6 Sol is highly capable at vulnerability research, remediation, debugging, and defensive testing. OpenAI has also strengthened safeguards intended to prevent misuse.

Those safeguards include built-in refusal behavior, real-time checks during response generation, account-level signals, usage monitoring, and tiered access controls.

For developers and administrators, this can make it easier to identify weaknesses, develop fixes, and strengthen system security.

The tradeoff is that even legitimate defensive requests may be stopped or trigger additional checks. OpenAI notes that its safeguards may occasionally restrict valid requests during the preview period.

Greater capability does not automatically make a system secure. Any proposed fix should be tested in an isolated environment.

When making a request, clearly state both your purpose and your authorization. For example, ask the model to “review the configuration of a system I manage and remediate vulnerabilities,” rather than to “find a way to break into this system.”

At the outset, specify who owns the target system, the authorized scope, prohibited actions, and what kinds of output are permitted.

Conditions for Making Safe Requests

• Target systems you own or manage
• Clearly state that you have permission to conduct the verification
• Aim for defense and remediation
• Limit the scope of the verification
• Review the procedure before execution
• Do not test directly in a production environment
• Prepare a backup
• Have a human perform the final review

GPT-5.6 API Pricing

OpenAI’s announcement also lists GPT-5.6 API pricing, calculated per 1 million tokens.

Sol costs $5 per 1 million input tokens and $30 per 1 million output tokens. Terra costs $2.50 for input and $15 for output, while Luna costs $1 for input and $6 for output, at the same token volume.

These rates apply when using the API to develop your own apps or services and are separate from the monthly ChatGPT Plus subscription fee.

The three pricing tiers let you control costs based on the workload. You can reserve Sol for critical analysis and send high-volume, routine processing to Terra or Luna.

Sol’s output price is relatively high, so long responses and large workloads can become expensive. Hard-coding one model for every request may also lead to unnecessary costs.

The amount billed in Japanese yen will vary with exchange rates and taxes. API usage is also billed separately from ChatGPT subscriptions, so Plus subscribers still pay additional fees when they use the API.

OpenAI has also improved the predictability of prompt caching in GPT-5.6. Explicit cache boundaries and a minimum cache retention time of 30 minutes have been introduced, and a 90% discount continues to apply to cache retrievals.

For applications that repeatedly send the same long instructions or documents, use caching and design each request so that only the changing content must be processed again.

Key Points on API Pricing

• Sol: $5 per 1 million input tokens
• Sol: $30 per 1 million output tokens
• Terra: $2.50 per 1 million input tokens
• Terra: $15 per 1 million output tokens
• Luna: $1 per 1 million input tokens
• Luna: $6 per 1 million output tokens
• Cache writes typically cost 1.25 times the input fee
• Cache reads receive a 90% discount
• Supports a minimum cache retention time of 30 minutes
• ChatGPT’s monthly subscription fee and API fees are billed separately

Errors Can Still Occur with GPT-5.6

GPT-5.6 is highly capable, but it can still be wrong.

Even the top-tier Pro mode did not solve every problem in OpenAI’s published scientific evaluation. As complexity rises, so does the risk of misreading assumptions, overlooking sources, making calculation errors, or introducing bugs.

Compared with earlier models, it can tackle a broader range of difficult problems and help identify the parts that need human verification.

But detailed, polished answers can sound convincing even when they contain mistakes. A confident tone is not evidence of factual accuracy.

For medicine, law, finance, contracts, security, and production code changes, always verify the output against official sources or with a qualified expert.

A safer prompt is: “Separate verified facts, unknowns, assumptions, and verification methods,” rather than simply asking for the answer.

For important answers, also check the primary source, the document date, the verification procedure, and plausible counterarguments.

How to Verify Answers

• Check official websites and public documents
• Check the publication and update dates of the information
• Have the model distinguish between facts and speculation
• Verify the calculation process
• Run code in a test environment
• Check for differences before making changes
• Test whether the same result is obtained using a different method
• Consult an expert for important decisions

Practical Tips for Using GPT-5.6 Efficiently

The most effective way to use GPT-5.6 is to scale the thinking level with the difficulty of the task instead of defaulting to the highest setting.

Use “Instant” for simple questions, sentence rewrites, and short explanations. Choose “Medium” for research, code fixes, and comparisons involving several conditions, then move to “High” only when Medium is not enough.

For higher-risk tasks—such as reviewing an entire architecture, investigating a large codebase, or performing specialized data analysis—start by asking GPT-5.6 for a plan only.

This approach gives you GPT-5.6’s full capabilities when they matter without consuming more of your allowance than necessary. Separating planning, execution, and verification also makes faulty changes easier to catch.

Breaking work into smaller steps takes more effort than issuing a single command, but that extra effort provides valuable safeguards on critical tasks.

Broad instructions such as “Fix everything” may lead to unintended changes. Clearly define both the scope of the work and the areas that must remain untouched.

For development work, ask for an investigation first and review the proposed changes before authorizing implementation. For blog posts and documents, approve the structure before drafting the body to reduce later revisions.

Include five elements in the prompt: purpose, target, constraints, completion criteria, and verification method.

A Reliable Order for Making Requests

• Communicate the purpose of the task
• Specify the target files or documents
• Specify which parts must not be changed
• Initially, request only research and planning
• Review the identified issues and proposed changes
• Ask the AI to implement only the approved changes
• Ask the AI to test and verify its work
• Ask the AI to explain the changes at the end

Current Notes Regarding GPT-5.6

As of July 2026, GPT-5.6 is rolling out in phases, so it may reach eligible accounts at different times.

The available benchmarks, detailed performance comparisons, and long-term usage limits may also change or expand over time.

In its June 26, 2026 announcement, OpenAI said it would publish a broader set of evaluations when the model becomes generally available. The limited-preview results alone do not show that GPT-5.6 outperforms every other model in every use case.

A phased rollout lets OpenAI check for bugs and validate safeguards as access expands.

It also means that interfaces and options can vary between users, even when they subscribe to the same plan.

Do not rely solely on social media posts or unofficial comparisons when judging performance or usage limits. Check your own interface, the official Help Center, and OpenAI’s release notes.

Where official information is not yet available, avoid filling the gaps with speculation and wait for OpenAI to publish further details.

To see what is available to you, check the model selector, your subscription plan, and the latest update date shown in the official Help Center.

Matters That Cannot Be Determined at This Time

• The completion date for rollout to all users
• Details of future changes to usage limits
• Features to be added to each plan in the future
• Practical performance differences during long-term use
• Detailed comparisons with GPT-5.5 across all fields
• Future pricing revisions
• Future changes to the model selection screen
• Details of additional evaluations to be released after general availability

Summary

GPT-5.6 is more than an update for faster answers to simple questions. It is a model family built to work through complex tasks in programming, specialized research, scientific analysis, and computer use.

The family has three tiers: the top-tier Sol, the balanced Terra, and the fast, low-cost Luna. Standard ChatGPT conversations primarily use Sol, while Terra and Luna are available in Work, Codex, and the OpenAI API.

ChatGPT Plus includes the Medium and High settings for GPT-5.6 Sol, but not Extra High or Pro. GPT-5.5 Instant still provides fast everyday responses, so GPT-5.6 has not replaced every existing model.

A practical setup is Instant for everyday questions, Medium for general research and development, and High for difficult problems. Starting with Medium and stepping up only when necessary provides a better balance of speed and usage limits than choosing the highest setting immediately.

To get started, open the ChatGPT model selector and look for “Medium” and “High.” If they are available, try Medium for your everyday development and research, then compare the results with GPT-5.5 on your own tasks.

If you are unsure which setting to choose, remember that GPT-5.6 is most useful when you reserve deeper reasoning for the tasks that genuinely need it. As of July 2026, that is more practical than running every request at the highest setting.