OpenAI GPT-5.6 Sol Limited Preview: Model Tiers Shift to Sol, Terra, and Luna

What was announced?

OpenAI announced on June 26 that it was beginning a limited preview of the GPT-5.6 series. The new system has three branches. Sol was described as the flagship model, Terra as a balanced model for everyday tasks, and Luna as a fast, lower-cost model.

OpenAI said Terra delivers performance competitive with GPT-5.5 while costing twice less, and that Luna covers the lowest-cost range. In other words, this announcement is closer to a new model family that separates performance, speed, and cost, rather than a single new model.

Why a limited preview?

The important part is the scope of access. OpenAI said this preview will first be provided to selected trusted partners and organizations across API and Codex. It is preparing broader availability across ChatGPT, Codex, and API, but as of today, this is not a full public release.

OpenAI explained that, through prior discussions with the U.S. government, it shared model capabilities and release plans, and that at the government’s request, it is starting with a small number of partners whose participation has been shared with the government. TechCrunch also covered the background of this limited rollout in a separate report. So for now, it is more accurate to see this not as “GPT-5.6 that everyone can use immediately,” but as a phased rollout that also includes safety validation.

Where are the performance highlights?

OpenAI introduced GPT-5.6 Sol as its top-end model and said it showed improvements in coding, life sciences, and cybersecurity evaluations. In coding in particular, it highlighted results on Terminal-Bench 2.1, which evaluates terminal command work, planning, iteration, and tool coordination.

Sol also adds max reasoning effort, which allows the model to reason more deeply for longer, and OpenAI introduced an ultra mode that uses multiple sub-agents to handle complex tasks faster. However, OpenAI said broader evaluation results will be shared when the model becomes generally available, so the currently disclosed figures should be read as preview-stage information.

Cyber safety measures are attached too

One area OpenAI explained at length in this announcement was safety. GPT-5.6 Sol was described as stronger at finding vulnerabilities and defensive security work, but according to OpenAI, it did not cross the Cyber Critical threshold. In Chromium and Firefox evaluations, the model found bugs and exploit components, but under the test conditions, it did not autonomously build a full attack chain.

OpenAI also said it invested more than 700,000 A100-equivalent GPU hours into automated red-teaming. Its approach layers refusal training inside the model, real-time checks during generation, account-level signals, tiered access, monitoring, and enforcement. As a result, OpenAI noted that even legitimate security work may be delayed or blocked during the preview period.

Pricing and what to watch in actual use

The pricing OpenAI disclosed is based on 1 million tokens. Sol is $5 for input and $30 for output, Terra is $2.50 for input and $15 for output, and Luna is $1 for input and $6 for output.

Models after GPT-5.6 will also include more predictable prompt caching, such as explicit cache breakpoints and a 30-minute minimum cache lifetime. Cache writes are priced at 1.25 times the uncached input rate, while cache reads use a 90% discount structure. In the end, developers will need to calculate not only performance, but also model tier, cache design, and output token costs.

The conclusion as of today

GPT-5.6 Sol may look like a model performance story from its name alone, but the actual announcement is an update that bundles together release controls, safety validation, pricing, and Codex/API usage. General users will need to wait for the timing of a full ChatGPT rollout and the real-world difference it brings, while developers should first check API costs and cache policy after the limited preview.

So the most accurate reading right now is: “OpenAI has revealed its next-generation model, but it is tying stronger capabilities to a more cautious release and safety framework.”