Direct answer: OpenAI launched in 26 June 2026 a limited preview of GPT-5.6 with three models: Sun, Earth, and Moon. The family arrives first through the API and Codex for selected organizations, not yet available in ChatGPT during the preview, promising advances in reasoning, code, science, cybersecurity, and operational safety.
The official OpenAI release presents GPT-5.6 as a new family of models, not just a single model. The logic is simple: Sun is the most capable model, Earth is the balanced option for daily work, and Moon is the fastest and most affordable version.
In clear English: OpenAI is organizing its lineup in levels of capability, cost, and speed. This matters because companies don't buy "intelligence" in the abstract. They choose which model should execute which task, with what cost, risk, and limit.
At XMACNA, the interpretation is operational. A stronger model raises the ceiling of what an agent can do but does not replace process, supervision, recording, and metrics. A Digital Employee must continue serving, qualifying, recording, and escalating to a human when needed, even when the AI engine changes behind the scenes.
What did OpenAI announce?
OpenAI began the limited preview of the GPT-5.6 series with three names: Sun, Earth, and Moon. The announcement positions Sun as the flagship model, Earth as a balanced model for daily use, and Moon as the fast and lower-cost option.
According to OpenAI itself, Earth has competitive performance with GPT-5.5 at half the cost. Moon is the lowest-cost version in the family. Sun focuses on the promise of greater capacity, particularly in long tasks, programming, scientific research, and cybersecurity.
The company also states it intends to make GPT-5.6 Sun, Earth, and Moon more widely available in the coming weeks. For now, the initial release is restricted.
This makes the launch news have two layers. The first is technical: there is a new family of models. The second is strategic: OpenAI treats access, security, and capacity as part of the same package. The model arrives strong but not loose.
Who can access GPT-5.6 now?
During the preview, access is limited to a small group of partners and trusted organizations. The OpenAI support article clarifies that GPT-5.6 is available in the preview via API and Codex, subject to approval by each participating organization.
The important point: GPT-5.6 is not available in ChatGPT during this phase. OpenAI also informs that there is no public sign-up, open waitlist, or self-service program for individual consumers.
For companies, this avoids a common mistake: planning a critical operation based on a resource not yet widely provisioned. If the organization received an invitation, it must follow OpenAI's access flow. If not, the professional path is to monitor the release and prepare internal tests, not promise production before real availability.
Sun, Earth and Moon: what is the difference?
The new nomenclature separates generation and capability level. The number 5.6 identifies the model generation. Sun, Earth and Moon identify tiers that may evolve at different rates.
Sun is the top of the family. It is the model OpenAI describes as the most capable, focused on tasks requiring deeper reasoning, longer deliberation time, and coordination of complex work.
Earth is the mid-tier. The idea is to be strong enough for everyday work but with a lower cost than the main model. In business operations, this type of tier is usually the most interesting for volume: customer service, triage, summarization, classification, light analysis, and sales flow support.
Moon is the fast and economical option. In a company, such models make sense for high-frequency and lower-risk tasks: intent recognition, preparing short drafts, message classification, answering simple questions, and triggering service routes.
The takeaway for managers is clear: not every task deserves the most expensive model. Value appears when the company routes the right task to the right model.
What changes in capability?
The announcement highlights gains in programming, biology, and cybersecurity workflows. OpenAI cites performance on Terminal-Bench 2.1 for command-line tasks, GeneBench v1 for genomics and quantitative biology analyses, as well as ExploitBench and ExploitGym for security tasks.
These benchmarks should not be read as an automatic promise for any company. They indicate a direction: models are becoming more capable in long tasks, with more stages, tools, and decision-making. This is precisely the type of task that brings the model closer to real agents.
OpenAI also presents a new 'max' reasoning effort to give Sun more thinking time. Additionally, the announcement mentions an 'ultra' mode, which uses sub-agents to speed up complex work.
Translating to operations: the model is not being designed just to respond better. It is being designed to sustain longer, coordinated, and difficult tasks. This increases productivity potential and also raises the need for governance.
Safety: why is the release phased?
OpenAI dedicates a large part of the announcement and the system card for GPT-5.6 to safeguards. The company says Sun, Earth, and Moon come with their most robust safety stack so far, with special attention to cybersecurity, biology, sensitive requests, and repeated abuse.
The official text also explains the phased release logic. OpenAI states it is coordinating the preview with selected partners before opening broader access. In cybersecurity, the stated goal is to preserve legitimate uses like code review, defensive research, patch development, debugging, and education, while making prohibited offensive use harder.
For a company, this is more important than it seems. When models can perform more difficult tasks, the usage policy stops being a footnote. It becomes architecture. It’s necessary to define what the agent can do alone, what it can only suggest, and what requires human approval.
This is where AI agents stop being a prompt game and become operational design.
Price and caching: how much does GPT-5.6 cost?
In the Help Center, OpenAI reports prices per 1 million tokens: GPT-5.6 Sun costs US$ 5 per input and US$ 30 per output; GPT-5.6 Earth costs US$ 2,50 per input and US$ 15 per output; GPT-5.6 Moon costs US$ 1 per input and US$ 6 per output.
There is also a change in prompt caching. OpenAI reports support for explicit cache breakpoints, a minimum cache lifetime of 30 minutes, cache writing at 1,25x the input price without cache, and reading with a discount of 90% over cached input.
This is a technical detail but quickly becomes a business decision. An operation with many repeated messages, stable context, and long instructions can benefit from predictable caching. On the other hand, an operation that changes context every conversation needs to measure if the gain shows in cost per completed task.
The number that matters is not just price per token. It is the cost per task resolved.
What does this change for companies?
It raises the expectation bar. More capable models expand the type of work that can be delegated, especially when involving reasoning, tools, analysis, and coordination.
But it also makes it clearer that the company needs a layer between the model and the operation. Without this layer, every release becomes anxiety: swap model, change prompt, break flow, measure by feeling, and discover later that customer service looked good but the CRM remained empty.
A serious company needs to test GPT-5.6 or any new model with real cases: anonymized conversations, commercial objections, documents, handoff flows, safety rules, cost, latency, and record quality. In IA process automation, the question is not "which model won the benchmark?" It is "which model performs this function with less rework, less risk, and better results?".
For sales, for example, a SDR with AI needs to qualify, ask the minimum necessary, record objections, trigger the salesperson, and preserve context. If the new model answers better but loses funnel discipline, the company hasn’t advanced. It only refined the error.
How does XMACNA view this launch?
XMACNA sees GPT-5.6 as capacity infrastructure. It is important, newsworthy, and may open new possibilities. But value for the client appears when that capacity becomes an executed function.
A well-designed Digital Employee doesn’t depend on hype. It separates function, data, tools, limits, and supervision. The model behind can change, as long as the operation remains predictable.
In practice, this means:
- test models by function, not enthusiasm;
- measure cost per completed task, not just price per token;
- maintain fallback to previous models;
- preserve records in the Intelligent Dashboard;
- define when the human takes over;
- separate response, decision, and action;
- assess safeguards before automating sensitive tasks.
This is the point most companies will discover now. The stronger the model, the more important the process around it becomes.
In summary
- OpenAI announced a limited preview of GPT-5.6 with three models: Sun, Earth, and Moon.
- The preview starts with the API and Codex for selected organizations, not yet available in ChatGPT.
- Sun is the most capable model; Earth balances performance and cost; Moon prioritizes speed and price.
- The launch highlights programming, biology, cybersecurity, a new 'max' reasoning effort, 'ultra' mode, safeguards, and phased release.
- For companies, the decision is not "use the newest model." It is to know which function it executes best, at what cost, with what limits, and what evidence.
If your company wants to turn AI launches into real work, start with the XMACNA AI Assessment. Models change fast. A well-designed process keeps working.
Frequently asked questions
What is GPT-5.6 Sun?
GPT-5.6 Sun is the main model of the new GPT-5.6 family announced by OpenAI. It is presented as the most capable model in the series, targeted at tasks requiring deep reasoning, programming, science, cybersecurity, and long work with tools.
Is GPT-5.6 already available in ChatGPT?
Not during the initial preview. According to OpenAI’s Help Center, GPT-5.6 is available in preview via API and Codex for selected organizations. OpenAI states it intends to expand availability to ChatGPT, Codex, and API in the coming weeks.
What is the difference between GPT-5.6 Sun, Earth, and Moon?
Sun is the most capable tier, Earth is the balanced option for everyday work, and Moon is the fast and economical choice. The idea is to allow clearer choices among intelligence, speed, and cost.
How much does GPT-5.6 cost?
According to OpenAI’s Help Center, per 1 million tokens, Sun costs US$ 5 input and US$ 30 output; Earth costs US$ 2,50 input and US$ 15 output; Moon costs US$ 1 input and US$ 6 output. These prices apply to the preview reported by OpenAI.
What should companies do with this release?
Companies should treat GPT-5.6 as a new capability to be tested, not as an automatic swap. The professional path is to select specific functions, create tests with real cases, measure cost per task, quality, safety, latency, fallback, and operational impact before production.