OpenAI began the limited preview of the GPT-5.6 family on June 26, 2026. This is not a single model release, but a new model tiering system: Sol is the flagship model, Terra is the balanced model for daily work, and Luna focuses on speed and cost.
According to OpenAI, Terra can compete with GPT-5.5 at about half the cost, while Luna offers strong capability at a lower price. Sol is the focus of this preview. OpenAI describes it as its strongest model so far, paired with its most robust safety stack to date.
Limited preview first, broader access later
GPT-5.6 will not be fully open from day one. During preview, OpenAI says the models will first be available through API and Codex to selected trusted partners and organizations, then expand quickly to ChatGPT, Codex, and API users.
There is also a special release context. Before launch, OpenAI discussed the model plan and capabilities with the U.S. government. At the government’s request, OpenAI first opened access to a small group of trusted partners and shared information about those participants with the government. OpenAI also argues that this kind of government access process should not become the long-term default, because it delays access for developers, enterprises, cyber defenders, and global partners.
So the current status of GPT-5.6 is:
- Limited preview has started.
- Initial access is through API and Codex for selected trusted partners.
- Broader availability for ChatGPT, Codex, and API users is still rolling out.
- Final availability, rate limits, and default entry points require later announcements.
Sol, Terra, Luna: the new model tiers
GPT-5.6 introduces a clearer naming system. The number represents the model generation, while Sol, Terra, and Luna represent long-term capability tiers that can evolve at their own pace.
This is different from using one model name plus version numbers. It gives users a more direct selection framework:
| Model | Positioning | Published price |
|---|---|---|
| GPT-5.6 Sol | Flagship model, highest capability | $5 / million input tokens, $30 / million output tokens |
| GPT-5.6 Terra | Balanced model for daily work | $2.50 / million input tokens, $15 / million output tokens |
| GPT-5.6 Luna | Fast, low-cost model | $1 / million input tokens, $6 / million output tokens |
GPT-5.6 also introduces more predictable prompt caching: explicit cache breakpoints and a minimum 30-minute cache lifetime. For GPT-5.6 and later models, cache writes are billed at 1.25 times the uncached input price, while cache reads continue to receive a 90% cached-input discount.
New capabilities: max reasoning and ultra mode
Sol’s improvements concentrate on three areas: coding, biology, and cybersecurity.
OpenAI says GPT-5.6 introduces a new max reasoning effort that gives Sol more time for deep reasoning. It also adds a new ultra mode that uses sub-agents to go beyond a single-agent setup and accelerate complex tasks.
For developers, the most important area is coding and tool coordination. OpenAI says GPT-5.6 Sol reaches a new high on Terminal-Bench 2.1, a benchmark focused on command-line workflows that require planning, iteration, and tool coordination, similar to products such as Codex.
In biology, OpenAI mentions GeneBench v1. In long-horizon genomics and quantitative biology analysis, Sol uses fewer tokens than GPT-5.5 while achieving stronger results.
Stronger cybersecurity capability, heavier safeguards
The safety section of the announcement is substantial. OpenAI states that GPT-5.6 Sol is its strongest model for cybersecurity so far, improving performance and efficiency in long-horizon security tasks, including vulnerability research and exploit-related work.
But OpenAI’s conclusion is not “open more attack capability.” It emphasizes that the model is better at helping users find and fix vulnerabilities, and is not reliable at executing end-to-end attacks. In Chromium and Firefox evaluations, Sol identified vulnerabilities and exploit primitives, but did not autonomously generate a full usable attack chain under test conditions.
OpenAI also says GPT-5.6 Sol does not cross the Critical cybersecurity threshold in its Preparedness Framework. Even so, OpenAI uses a staged release because benchmarks cannot cover every way a model may be combined and used.
How the layered safety stack works
GPT-5.6 safety is not a single refusal rule, but a layered set of safeguards:
- Safety behavior trained into the model to refuse prohibited cybersecurity assistance.
- Real-time classifiers for cybersecurity and biological misuse during generation.
- Pausing high-risk generations and sending them to a larger reasoning model for review.
- Account-level risk signals and related conversation review.
- Differential access, monitoring, enforcement, and ongoing testing.
The cost is that legitimate users may see false blocks or delays during preview. In dual-use areas such as cybersecurity, defensive testing and offensive activity can look similar in early requests. OpenAI says one preview goal is to collect feedback and reduce unnecessary blocking and latency.
Large-scale automated red teaming
OpenAI also disclosed significant automated red-team investment: more than 700,000 A100-equivalent GPU hours to find general jailbreaks. These are not narrow attacks against a single scenario, but potentially general failures across prompts and contexts.
This shows frontier model safety testing moving from a few human attack examples toward large-scale automated search and continuous evaluation. OpenAI will also combine third-party expert red teaming, reproduce new issues, evaluate and prioritize them, fix them, and add them to future evaluation sets.
High-speed version on Cerebras
OpenAI also says GPT-5.6 Sol will arrive on Cerebras in July, reaching up to 750 tokens per second. Initial access will be limited to selected customers and expand with capacity.
This matters because it puts the highest-capability model and high output speed on the same path. For coding agents, long-document processing, and interactive analysis, inference speed directly affects product experience.
What signal this release sends
The limited preview of GPT-5.6 Sol shows OpenAI moving on three fronts:
- Establishing clearer tiers with Sol, Terra, and Luna.
- Putting stronger Agent, coding, biology, and cybersecurity capability into the new generation.
- Using more cautious staged release and more complex safety protections for high-capability models.
Ordinary users do not need to migrate immediately because GPT-5.6 is not broadly available yet. Developers and enterprises should watch three changes early: new model pricing, prompt caching rules, and the impact of max reasoning effort and ultra mode on Agent workflows.
If OpenAI expands availability as planned, GPT-5.6 may become the new main model line after GPT-5.5. Sol handles highest capability, Terra handles daily cost-performance, and Luna handles cheaper, faster workloads. This tiering is more useful for real product selection than chasing a single “best model” name.
How developers should prepare
GPT-5.6 Sol is still in limited preview, so ordinary developers may not be able to use it immediately. But migration preparation can start now. The most practical step is to build an evaluation set from high-value GPT-5.5 or current-model tasks: complex code changes, tool use, long-context analysis, structured output, refusal boundaries, and cost-sensitive work.
When Sol, Terra, and Luna become more broadly available, do not test only Sol. Sol is for the hardest tasks, but the volume in real products usually runs on Terra or Luna. The point of tiering is to place each task in the right capability tier, not to send every request to the flagship model.
For Codex users, two things deserve attention: whether Sol is more stable on multi-file changes, test-failure repair, and repository understanding; and whether ultra mode plus max reasoning effort is worth the extra cost. Higher reasoning cost makes sense only when it significantly reduces human rework.
Selection approach
Think of the three tiers as a pipeline: Luna for classification, cleaning, light summaries, and high throughput; Terra for default conversation, ordinary coding, and content generation; Sol for complex Agents, hard problems, and failure retries.
A mature integration should not hard-code one model. It should route by task type, risk level, context length, and failure count. That is how you capture both new capability and cost advantages.
Original: Previewing GPT-5.6 Sol: A new generation of models