GPT-5.6 Sol (2026): Pricing, API, Benchmarks, Specs & How to Use It

GPT-5.6 is OpenAI’s current GPT model family. GPT-5.6 Sol is the flagship model, built for harder coding, research, computer use, design, and professional work. Terra is the lower-cost middle tier, while Luna is the fastest and cheapest model in the family, according to OpenAI’s GPT-5.6 overview.

That doesn’t make Sol the right choice for every task. If you’re choosing a model for an API workload, coding agent, internal tool, or other multi-step workflow, the real question is whether Sol’s extra capability improves the result enough to justify the cost.

Pricing and access checked: August 12, 2026.

These specs come from OpenAI’s current GPT-5.6 Sol model documentation.

What is GPT-5.6 Sol?

GPT-5.6 is the model generation. Sol is the highest-capability model in that generation.

OpenAI splits the family into three tiers:

• Sol: The flagship model for complex work.
• Terra: A lower-cost option for everyday professional tasks.
• Luna: The fastest and least expensive GPT-5.6 model.

OpenAI says Sol, Terra, and Luna are intended to be durable capability tiers rather than temporary names attached to one release.

There’s also an API naming detail worth knowing. If you call gpt-5.6, OpenAI currently routes that request to GPT-5.6 Sol. You can also call gpt-5.6-sol directly.

Where Sol becomes interesting is the amount of work it can carry through. A simple extraction task may not need it. A job that involves inspecting files, using tools, fixing errors, checking the result, and continuing through several steps is a much better fit.

That distinction matters for multi-step agentic workflows, where model quality is only one part of whether the full workflow succeeds.

When was GPT-5.6 released?

GPT-5.6 became generally available on July 9, 2026. OpenAI had started a limited preview of Sol, Terra, and Luna on June 26, as detailed in its GPT-5.6 Sol preview announcement.

A few later updates matter because they changed pricing and ChatGPT access:

• June 26: OpenAI began the limited GPT-5.6 preview.
• July 9: Sol, Terra, and Luna reached general availability.
• July 30: OpenAI cut Terra and Luna API prices and introduced Fast mode for Sol.
• August 6: OpenAI updated Sol in ChatGPT and expanded Luna access for Free and Go users.

If you’re comparing GPT-5.6 prices across older reviews, check the date. Terra and Luna are now much cheaper than they were at launch.

What’s new in GPT-5.6 Sol?

The biggest change isn’t a larger context window. GPT-5.5’s model documentation already lists a 1,050,000-token context window and 128,000 output tokens, the same headline limits listed for Sol.

The more useful gains show up in coding, tool use, computer interaction, design, and long-running work.

Stronger coding across longer tasks

Sol is aimed at coding work that takes more than one clean prompt.

Think of a task where the model needs to understand an unfamiliar repository, decide which files matter, make several changes, run tests, inspect failures, and keep working.

In OpenAI’s GPT-5.6 benchmark results, Sol with max reasoning scores 80 on the Artificial Analysis Coding Agent Index, compared with 76.4 for GPT-5.5.

Independent testing adds some useful caution.

In Sonar’s 4,444-task Java evaluation, Sol passed 81.99% of tasks, compared with 78.66% for GPT-5.5. Sonar also found higher bug and vulnerability densities in Sol’s generated code.

For engineering teams, that’s more useful than simply saying Sol is “better at coding.” It completed more tasks correctly in Sonar’s test, but strong automated testing and code review still matter.

Better coordination across tools

GPT-5.6 supports Programmatic Tool Calling in the Responses API.

OpenAI says the model can write small programs that coordinate tools, process intermediate results, and keep only the information needed for the next step. That can reduce model round trips and wasted tokens in tool-heavy work.

This is useful when one task spans several systems. A research workflow may need to search, inspect documents, filter results, run code, and then decide what evidence is worth keeping.

The model can handle part of that coordination, but your application still decides which tools it can access and what it’s allowed to do.

Better browsing and computer use

GPT-5.6 also improved on evaluations where the model has to work through websites or software rather than only answer a prompt.

OpenAI reports:

• 90.4% on BrowseComp for standard Sol, versus 84.4% for GPT-5.5.
• 62.6% on OSWorld 2.0, versus 47.5% for GPT-5.5.
• 92.2% on BrowseComp with the multi-agent ultra setup.

The 92.2% result shouldn’t be treated as the score for ordinary Sol. It comes from the ultra configuration, which uses several agents.

These results matter more if you’re building software that needs to browse, use tools, or interact with interfaces. They tell you much less about a simple writing or classification task.

Better finished documents and presentations

OpenAI is also pushing Sol beyond plain text output.

The company says GPT-5.6 is better at creating editable presentations, following reference templates, handling spreadsheets and financial models, and keeping document layouts consistent.

Artificial Analysis’ GPT-5.6 evaluation found a similar pattern. In its AA-Briefcase knowledge-work evaluation, Sol with max had the highest Presentation Elo, while another model scored higher on its measure of analytical quality.

That difference is worth paying attention to. A polished presentation can still contain weak analysis. If the numbers matter, check the reasoning as well as the layout.

What are GPT-5.6 max, ultra, and Sol Pro?

These names are easy to mix up because they describe different things.

What is GPT-5.6 max?

max is a reasoning-effort setting.

OpenAI says it gives GPT-5.6 more time than xhigh to explore alternatives, run checks, and revise its approach. That can help on hard problems, but it also means spending more compute on the task.

You don’t need max because a task sounds important. Use it when your own tests show that the extra reasoning changes the result.

What is GPT-5.6 ultra?

ultra goes beyond one agent.

OpenAI says its default setup coordinates four agents in parallel. That lets separate agents work on parts of a problem at the same time before their work is combined.

The tradeoff is higher token use.

So ultra isn’t simply another step above max. One controls reasoning effort. The other adds parallel agents.

What is GPT-5.6 Sol Pro?

Sol Pro is the highest-capability GPT-5.6 option offered in ChatGPT for eligible plans.

According to OpenAI’s current ChatGPT model access documentation, the current setup uses:

• Medium: GPT-5.6 Sol
• High: GPT-5.6 Sol
• Extra High: GPT-5.6 Sol
• Pro: GPT-5.6 Sol Pro

Plus currently includes Medium and High. Pro, Business, and Enterprise include the broader Sol options, although managed workspace settings can affect availability.

What is the GPT-5.6 Sol context window?

GPT-5.6 Sol has a 1,050,000-token context window and can generate up to 128,000 output tokens, based on OpenAI’s current model limits.

That gives you room for large codebases, long documents, extended sessions, and workflows with a lot of source material.

But a million-token window doesn’t mean you should fill it on every request.

Requests above 272K tokens cost more

OpenAI applies higher pricing when a Sol request contains more than 272,000 input tokens.

Once you cross that threshold, the full request is charged at 2x the normal input rate and 1.5x the normal output rate. It isn’t only the tokens above 272K that get the higher price.

For a production workload, that can change the economics quickly.

A large context window reduces the pressure to split everything into tiny pieces. It doesn’t remove the value of good retrieval, filtering, caching, and sending the model only the information it needs.

More context doesn’t guarantee perfect recall

OpenAI’s own long-context results show this clearly.

On its MRCR v2 eight-needle evaluation, Sol scored 91.5% between 256K and 512K tokens. Between 512K and 1M tokens, the score fell to 73.8%. GPT-5.5 scored 74.0% in that larger range.

So Sol can accept a million tokens, but that isn’t the same as reliably finding every important detail inside a million-token prompt.

If your product depends on long documents or large knowledge bases, test recall on the documents your users will actually send.

How much does GPT-5.6 Sol cost?

GPT-5.6 Sol currently costs $5 per million input tokens and $30 per million output tokens. Cached input costs $0.50 per million tokens, according to OpenAI’s Sol model pricing.

The current GPT-5.6 family prices are:

Sol: $5 input / $30 output per 1M tokens.

Terra: $2 input / $12 output per 1M tokens.

Luna: $0.20 input / $1.20 output per 1M tokens.

OpenAI’s July 30 pricing update cut Terra to $2/$12 and Luna to $0.20/$1.20. Sol stayed at $5/$30.

Sol cache writes are charged at 1.25x the uncached input rate. Cache reads receive the lower cached-input price.

Token price isn’t the same as task cost

Suppose Luna costs less per token but needs three attempts to complete a task that Sol gets right once.

Or Sol spends much longer reasoning on work that Terra already handles reliably.

Looking only at token price can hide both cases.

Artificial Analysis measures cost per task for this reason. Its July evaluation found Sol with max more expensive per task than Terra and Luna, while also scoring higher on its Intelligence and Coding Agent indexes.

For a real application, track:

• Successful completions
• Retries
• Total tokens
• Latency
• Tool failures
• Human review time
• Cost per successful result

Those numbers will tell you more than a model price page.

How do you use the GPT-5.6 Sol API?

For the OpenAI API, use:

gpt-5.6-sol

You can also use:

gpt-5.6

which currently routes to Sol.

OpenAI’s GPT-5.6 Sol API documentation lists text and image input with text output. Audio and video aren’t supported modalities for this model. Streaming, function calling, and structured outputs are supported, while fine-tuning isn’t currently available.

With the Responses API, Sol can use tools including:

• Web search
• File search
• Code interpreter
• Hosted shell
• Computer use
• MCP
• Image generation
• Skills
• Tool search

That makes the API more useful for applications where the model has to do something with information, not just generate a response.

For example, an agent might search a knowledge source, check a file, run code against the result, and decide what action should happen next.

GPT-5.6 Sol Fast mode

If latency matters, OpenAI also offers Fast mode for Sol.

OpenAI’s price-performance update says Fast mode can deliver up to 2.5x faster performance than Standard processing at twice the price, with no change to the underlying model intelligence. It replaced Priority Processing, while existing priority requests remain compatible.

That can make sense for an interactive product where waiting has a real cost. It’s harder to justify for background work that can finish a little later.

Is GPT-5.6 Sol free?

No. Free and Go ChatGPT users currently receive GPT-5.6 Luna, not GPT-5.6 Sol.

OpenAI’s current ChatGPT plan mapping lists:

Free and Go: Luna is the default GPT-5.6 model. Think also uses Luna.

Plus: Medium and High use GPT-5.6 Sol.

Pro: Medium, High, Extra High, and Sol Pro are available.

Business: Medium, High, Extra High, and Sol Pro are available.

Enterprise: The same Sol reasoning options are available, subject to workspace controls and rollout.

One detail can be confusing: Instant is currently powered by GPT-5.5 Instant. On eligible paid plans, ChatGPT can automatically switch a harder request from Instant to Medium, which uses GPT-5.6 Sol.

Terra and Luna aren’t selectable in standard paid ChatGPT conversations. They are available in other OpenAI products depending on your plan, including Work, Codex, and the API.

GPT-5.6 benchmarks

A benchmark is useful when it resembles the work you care about.

A coding score won’t tell you how well Sol handles customer conversations. A browsing score won’t tell you whether it will remember one sentence buried in a huge policy document.

A smaller set of results is more useful than a page full of scores.

Coding

GPT-5.6 Sol with max scored 80 on the Artificial Analysis Coding Agent Index. GPT-5.5 scored 76.4 in OpenAI’s published comparison.

Artificial Analysis independently reports the same 80-point Coding Agent Index result for Sol with max. It also scored Sol at about 59 on its broader Intelligence Index during its July evaluation.

Sonar gives a different type of evidence. Sol passed 81.99% of its 4,444 Java tasks, ahead of GPT-5.5 at 78.66%, but produced more static-analysis findings in several bug and security categories.

If you plan to use Sol for production code, the second result matters as much as the first.

Browsing and computer use

Standard Sol scored 90.4% on BrowseComp and 62.6% on OSWorld 2.0 in OpenAI’s evaluation.

GPT-5.5 scored 84.4% and 47.5% on those tests.

Those gains make Sol more interesting for agents that browse or work inside software. They don’t prove it will complete every real website or computer task successfully.

Long context

Sol scored 91.5% on OpenAI’s 256K to 512K MRCR test, compared with 81.5% for GPT-5.5.

In the harder 512K to 1M range, Sol scored 73.8%, and GPT-5.5 scored 74.0%.

Sol doesn’t beat GPT-5.5 on every evaluation. That’s one reason workload testing matters.

Professional work

Artificial Analysis placed Sol with max second in its AA-Briefcase knowledge-work evaluation and gave it the highest Presentation Elo in that test. Its analytical-quality result was weaker than the leading model.

If you need a model to produce a client-ready deck or spreadsheet, that is useful evidence. If the deliverable contains important calculations or recommendations, you still need to check the work behind the presentation.

GPT-5.6 Sol vs GPT-5.5

The difference between GPT-5.6 Sol and GPT-5.5 isn’t simply a bigger context window or a higher API price.

They currently share several major API specs:

• 1,050,000-token context window
• 128,000-token maximum output
• $5 per million standard input tokens
• $30 per million output tokens
• $0.50 per million cached input tokens

Sol has a newer knowledge cutoff, February 16, 2026, compared with December 1, 2025 for GPT-5.5.

The stronger reason to move comes from the workload.

Sol performs better than GPT-5.5 on several coding-agent, browsing, computer-use, tool-use, and professional-work evaluations. It also adds max, multi-agent capabilities behind ultra, and Programmatic Tool Calling as part of the GPT-5.6 generation.

But “newer” isn’t a migration plan.

If GPT-5.5 already handles a narrow extraction task reliably, moving it to Sol may make little difference. If your failures happen during long coding jobs, tool coordination, or computer interaction, Sol is more worth testing.

GPT-5.6 Sol vs Terra vs Luna

Don’t choose between the three models by asking which one is best overall. Choose based on the job.

Choose Sol when getting the answer wrong is expensive

Sol makes sense when more reasoning or better tool use can change whether the work succeeds.

Large code changes, complex debugging, multi-source research, and difficult computer-use tasks are good examples.

A simple text classification probably isn’t.

Choose Terra when you want a middle ground

Terra is the balanced tier.

At $2 input and $12 output per million tokens, it costs much less than Sol. OpenAI positions it for everyday work where you still need solid capability.

Test Terra when Luna misses too many edge cases, but Sol doesn’t improve your results enough to pay more for every request.

Choose Luna when scale matters

Luna costs $0.20 per million input tokens and $1.20 per million output tokens. That makes it far easier to use at high volume.

Well-defined extraction, classification, routing, and background tasks are natural places to test it.

The tradeoff becomes clearer on harder work. In OpenAI’s 512K to 1M MRCR evaluation, Luna scored 41.3%, compared with 73.8% for Sol and 72.5% for Terra.

A useful rule is simple: use the least expensive model that reliably passes your own tests.

What is GPT-5.6 Sol best for?

Sol earns its higher price when the work is difficult enough that better reasoning or persistence saves real time.

Large software changes

You probably don’t need Sol to write a small utility function.

It becomes more useful when a coding task spans a repository, needs several edits, requires tests, and may need the model to recover from failures along the way.

Sonar’s results suggest Sol can complete more coding tasks than GPT-5.5, but the same evaluation also reinforces the need for automated code checks.

Research with several steps

Sol is a better fit when research means more than searching for one fact.

A harder workflow may involve browsing several sources, inspecting documents, extracting data, comparing evidence, and deciding what belongs in the final answer.

Its BrowseComp result and Programmatic Tool Calling support make that a more natural fit than a simple summary request.

Tool-heavy agents

If the model has to call tools, inspect the response, change course, and continue, Sol is worth testing.

The model still shouldn’t control every available system by default. Give it only the permissions needed for the job.

Computer-use workflows

Sol’s OSWorld improvement makes it relevant when an agent needs to operate software rather than only call an API.

Keep approval steps around destructive or sensitive actions. A better computer-use benchmark doesn’t make unrestricted access safe.

Presentations, spreadsheets, and polished deliverables

Sol is also a good candidate when the final output has to be usable by another person.

OpenAI highlights presentations, spreadsheets, documents, and frontend work as strengths, while Artificial Analysis found strong presentation quality in its own testing.

Polish is useful. It isn’t proof that every fact or calculation is right.

GPT-5.6 Sol limitations: what still needs human review?

Sol can do more on its own, but that makes good controls more important, not less.

Large context isn’t perfect memory: Sol accepts more than one million context tokens, but recall drops on some very long-context tests. Test the documents and retrieval patterns your product will use.

More reasoning costs time and tokens: max spends more compute, while ultra coordinates several agents. Use the extra capability when it changes the outcome, not as a default.

Generated code still needs verification: Sonar found a higher pass rate for Sol than GPT-5.5, but it also found more bug and security findings in several categories.

A capable agent still needs limits: Give it access only to the data and tools required for the task. Keep approval steps around actions that are costly, sensitive, or hard to reverse.

What shouldn’t you fully automate?

Keep a human or explicit approval rule involved when the action carries meaningful risk.

That can include:

• Production changes that could cause downtime
• Large payments, refunds, or purchases
• Account ownership or permission changes
• Requests that fall outside an approved policy
• Decisions based on conflicting or missing information
• High-stakes legal, medical, financial, or compliance decisions

For coding, human review doesn’t have to mean reading every line manually. Tests, static analysis, permissions, and deployment checks can handle part of the review.

For customer-facing agents, escalation matters when the agent doesn’t have enough information, a customer needs an exception, or an action falls outside the workflow it has been approved to handle. Chatbase’s customer support AI agent breakdown goes deeper into where that line should sit.

How GPT-5.6 fits into a multi-model AI agent

A model is only one layer of an AI agent.

The model handles language and reasoning. The rest of the system decides what knowledge it can use, which tools it can call, what actions it can take, and when a person should step in.

That’s especially important in AI customer support. An order-status question may be cheap and predictable. A billing exception involving account history and several business rules is a different problem.

Chatbase currently supports GPT-5.6 Terra and GPT-5.6 Luna for AI agents. Its GPT-5.6 Terra and Luna update positions Luna around speed and high-volume support, while Terra is aimed at conversations that need stronger reasoning and instruction following. The update doesn’t list Sol, so Sol shouldn’t be presented as available in Chatbase yet.

This is also why using one model everywhere is rarely the most useful architecture.

A lower-cost model can handle repetitive work. A stronger model can take the cases where reasoning matters more. The agent around those models still needs the right data, rules, actions, and human handoff.

That is the layer that turns a capable model into something useful for customers.

To put that model choice into practice, you can create a Chatbase account and test GPT-5.6 Terra or Luna with your own support knowledge, workflows, and handoff rules.

FAQs

Is GPT-5.6 the same as GPT-5.6 Sol?

No. GPT-5.6 is the model family. Sol is its flagship model, alongside Terra and Luna. In the OpenAI API, the gpt-5.6 alias currently routes to GPT-5.6 Sol.

Is GPT-5.6 Sol free?

No. Free and Go ChatGPT users currently receive GPT-5.6 Luna rather than Sol. Sol is available on eligible paid ChatGPT plans and through the paid OpenAI API.

What is the GPT-5.6 Sol context window?

GPT-5.6 Sol has a 1,050,000-token context window and supports up to 128,000 output tokens. API requests above 272K input tokens use higher long-context pricing for the full request.

GPT-5.6 Sol vs Terra vs Luna: which should you use?

Use Sol when harder reasoning, coding, or tool use changes the outcome. Test Terra when you want a better balance between capability and cost. Luna makes more sense for high-volume, well-defined work where low cost matters most.

Should you use GPT-5.6 Sol?

GPT-5.6 Sol is most compelling when your problem is finishing difficult work reliably, not getting a quick answer.

Its clearest gains show up in coding agents, browsing, computer use, tool-heavy tasks, and professional deliverables. OpenAI also gives you ways to spend more compute when a hard problem warrants it through max, ultra, and Sol Pro.

But there’s little value in paying flagship prices for work a cheaper model already handles well.

Sol and GPT-5.5 currently share the same standard API price and context-window size. Terra costs much less. Luna costs less again. OpenAI’s own results also include tests where Sol barely improves on GPT-5.5 or doesn’t lead it.

Start with your workload. Measure whether the model finishes it correctly, how often it retries, how long it takes, and how much human review is still needed.

Then choose the cheapest model that meets the quality bar consistently.