# OpenAI GPT-5.6 Sol Launches in Limited Preview

> Source: [https://icharles.com/articles/openai-gpt-56-sol-limited-preview](https://icharles.com/articles/openai-gpt-56-sol-limited-preview) (canonical)
> Author: iCharles News — iCharles, https://icharles.com
> Published: 2026-07-08

## TL;DR

OpenAI launched GPT-5.6 Sol, Terra, and Luna on June 26, 2026, but only to a small group of trusted partners. The Trump administration asked for the restricted rollout under a voluntary framework that allows up to 30 days of federal review before broad release. OpenAI complied but said this process "should not become the long-term default." Broader availability across ChatGPT, Codex, and the API is planned for the coming weeks.

## What is GPT-5.6 Sol, and what just launched?

**GPT-5.6 Sol** is OpenAI's newest flagship model, released in limited preview on June 26, 2026. It sits at the top of the GPT-5.6 series, which also includes Terra, a balanced everyday model, and Luna, a fast, low-cost option. All three launched simultaneously — but only to a small group of trusted partners, not the general public.

OpenAI [confirmed the restricted rollout on its blog](https://openai.com/index/previewing-gpt-5-6-sol/), saying partners' participation has been shared with the U.S. government. Broader access via ChatGPT, Codex, and the API is planned for the coming weeks.

## Why did the U.S. government restrict the GPT-5.6 release?

The Trump administration asked OpenAI to limit the launch. The request follows President Trump's June executive order, which created a voluntary framework for AI developers to submit "covered frontier models" for up to 30 days of federal review before broader release to trusted partners.

The Commerce Department's Center for AI Standards and Innovation (CAISI) ran the pre-release testing. OpenAI sent technical experts to Washington and kept them available to answer questions throughout the evaluation, [according to Axios reporting cited by AI Weekly](https://aiweekly.co/alerts/commerce-clears-openais-gpt-56-for-broad-thursday-launch).

OpenAI is not the only company affected. Anthropic went through a parallel process for its Mythos and Fable models. The administration also ordered Anthropic to remove access to Fable 5 for any foreign national, which led Anthropic to take the model down entirely.

## What did OpenAI say about the government review process?

OpenAI complied with the request but made its position clear. "We don't believe this kind of government access process should become the long-term default," the company wrote in its June 26 blog post. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

OpenAI called the preview a "short-term step." The company said it is working with the administration to develop a cybersecurity executive order framework and "a repeatable process for future model releases."

Here's what we know so far: this is the first live test of the June executive order, and it has already produced a working precedent — one that Anthropic also experienced with its own models.

## What are GPT-5.6 Sol's key capabilities?

OpenAI says Sol is its strongest model yet, with improved agentic capabilities in coding, biology, and cybersecurity. Two new modes ship with Sol:

- **Max reasoning effort** — gives Sol more time to reason deeply on complex tasks
- **Ultra mode** — uses coordinated subagents to tackle highly complex work

On **Terminal-Bench 2.1**, which tests command-line workflows requiring planning, iteration, and tool coordination, Sol sets a new state of the art. On **GeneBench v1**, which evaluates long-horizon genomics and quantitative biology, Sol outperforms GPT-5.5 while using fewer tokens. On **ExploitBench**, Sol is competitive with Anthropic's Mythos Preview while using only about one-third of the output tokens.

OpenAI also notes Sol does not cross the "Cyber Critical" threshold under its Preparedness Framework. In tests involving Chromium and Firefox, Sol identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit.

## How does GPT-5.6 Sol handle cybersecurity safety?

OpenAI built safety guardrails directly into the model's core behavior rather than layering a separate filter on top. The company says this approach avoids the problem Anthropic encountered with Fable 5, where a classifier detecting high-risk topics would silently route requests to an older model — leading to false positives and user backlash.

GPT-5.6's layered safeguard stack includes:

- Model-level training to refuse prohibited cyber assistance, including jailbreak attempts
- Real-time classifiers that evaluate output as it is generated; for higher-risk cases, generation can be paused while a larger reasoning model reviews the conversation
- Account-level review across conversations to distinguish persistent malicious behavior from legitimate security work
- Differentiated access that preserves defensive use cases without making sensitive capabilities broadly available by default

OpenAI says Sol is better at helping users find and fix vulnerabilities than at carrying out end-to-end attacks.

## What is GPT-5.6 pricing?

The three models come with tiered pricing. OpenAI has also improved prompt caching to make repeated prompts cheaper and more predictable.

| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|-------|-----------------------|------------------------|
| Sol   | $5                    | $30                    |
| Terra | $2.50                 | $15                    |
| Luna  | $1                    | $6                     |

Terra offers competitive performance to GPT-5.5 at 2x lower cost. Luna is positioned as the lowest-cost option in the series.

For context on how OpenAI has approached [inference cost reduction](/articles/openai-inference-cost-halved-optimization) in recent months, the Luna pricing continues that trend toward cheaper, faster access.

## Who else is affected by the federal review framework?

The framework is not limited to OpenAI. Anthropic's Mythos and Fable models went through a parallel review process. The restriction on Fable 5 was reportedly lifted the following week after it was initially pulled. [TechCrunch reports](https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/) that Dean Ball, a former White House AI adviser and incoming OpenAI employee, has called the executive order framework a "de facto involuntary licensing regime."

Ball argues that without clearly defined safety standards, the process could cause endless launch delays — a concern relevant to anyone tracking [government AI partnerships](/articles/california-anthropic-claude-state-partnership) at the state and federal level.

The Anthropic situation also raised questions about [Claude's user growth](/articles/claude-paid-users-growth-2026) and what access restrictions mean for commercial momentum. The same dynamics now apply to OpenAI's GPT-5.6 rollout.

Developers watching the [Claude Cowork mobile launch](/articles/claude-cowork-mobile-web-launch) and other frontier model releases should note that federal pre-release review has now moved from theory to a documented, repeatable process.

OpenAI plans to make GPT-5.6 Sol, Terra, and Luna broadly available in the coming weeks, as it continues coordinating with the administration on the cybersecurity executive order framework.

## Frequently asked questions

****What is GPT-5.6 Sol and when did it launch?****

GPT-5.6 Sol is OpenAI's newest flagship model, launched in limited preview on June 26, 2026. It is the most powerful model in the GPT-5.6 series, which also includes Terra, a balanced everyday model, and Luna, a fast, low-cost option. All three are currently available only to a small group of trusted partners, with broader access planned for the coming weeks.

****Why is GPT-5.6 only available to trusted partners right now?****

The Trump administration asked OpenAI to restrict the rollout under a June executive order that allows up to 30 days of federal review for covered frontier models before broad release. The Commerce Department's Center for AI Standards and Innovation conducted pre-release testing. OpenAI complied but stated it does not believe this process should become the long-term default for model releases.

****How much does GPT-5.6 Sol cost?****

GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output tokens. Terra costs half that amount. Luna, the lowest-cost option, is priced at $1 per million input tokens and $6 per million output tokens. OpenAI also improved prompt caching to make repeated prompts cheaper and more predictable across all three models.

****How does GPT-5.6 Sol compare to Anthropic's Claude Mythos?****

On ExploitBench, GPT-5.6 Sol is competitive with Anthropic's Mythos Preview while using only about one-third of the output tokens. OpenAI also says Sol sets a new state of the art on Terminal-Bench 2.1 for coding workflows and outperforms GPT-5.5 on GeneBench v1 for biology tasks, using fewer tokens than its predecessor.

****What safety measures does GPT-5.6 Sol include?****

Sol uses a layered safeguard stack built directly into the model's core behavior. This includes model-level training to refuse prohibited cyber assistance, real-time classifiers that can pause generation for review, and account-level monitoring across conversations. OpenAI says Sol is better at finding and fixing vulnerabilities than carrying out end-to-end attacks, and it does not cross the Cyber Critical threshold under OpenAI's Preparedness Framework.