OpenAI Introduces GPT-5.6 with Sol, Terra, and Luna Variants — Currently Exclusive to Select Preview Partners, According to U.S. Government Reports

OpenAI has unveiled a limited preview of its latest AI model series, GPT-5.6, which includes three distinct variants: Sol, Terra, and Luna.

Sol is tailored for advanced tasks such as intricate coding and security research; Terra is designed for high-demand business functions, including customer support and document analysis; and Luna is optimized for speedy, cost-effective routine operations like drafting and summarization. Both Sol and Terra achieve remarkable benchmark scores, while Luna operates close to the performance of GPT-5.5 across several assessments despite being marketed as the fastest and most economical model within the family.

Initially, these models will be accessible to a select group of around 20 organizations, following OpenAI’s shared plans with the U.S. government. A general release is anticipated in “a few weeks.”

This phased launch correlates with an executive order from President Donald J. Trump that urges federal agencies to collaborate on assessing and validating the capabilities of emerging AI models to ensure their safety before widespread deployment.

As this review process is currently in motion, OpenAI has mentioned in its blog that it has outlined its plans and capabilities before the official launch date, stating, “at the government’s request, we are starting with a limited preview for a select group of trusted partners.”

OpenAI’s strategy of limited previewing follows recent actions by the U.S. government, which placed an export control order on Anthropic, one of OpenAI’s main competitors. This was prompted by vulnerabilities discovered in Anthropic’s advanced model, Claude Fable 5, leading them to revoke access to several models previously offered for cybersecurity research.

As OpenAI aligns its release strategy with White House protocols for a broader public launch, enterprises must adapt to a new environment characterized by real-time safety measures, compliance requirements, and structured token systems.

Comparing the three GPT-5.6 models: Sol, Terra, and Luna

The newly launched models are tailored to meet different enterprise needs and performance metrics.

Sol is positioned as the premium choice for high-stakes tasks that require deep reasoning, extensive coding, advanced workflows, and security-related applications. It commands the highest price point at $5.00 for every million input tokens and $30.00 for output tokens—matching GPT-5.5 rates, but offering significant performance enhancements for long-term coding and security tasks.

Terra strikes a balance between performance and affordability, aimed at large production settings that demand consistent results across substantial workloads without the complexity of the top-tier model, priced at $2.50 for inputs and $15.00 for outputs per million tokens.

Luna stands as the most economical choice, tailored for speed and everyday applications, best for routine activities and workflows where quick response and scalability take precedence over deep reasoning, and is priced the lowest at $1 for inputs and $6 for outputs per million tokens.

Industry insiders indicated that the new naming convention was crafted to distance itself from earlier “nano” and “mini” model designations, as these models are more aligned with varying use cases rather than differing significantly in size or intelligence. OpenAI highlighted that this new naming approach not only signifies the model generation but also identifies capability tiers, allowing users clearer options in terms of intelligence, speed, and costs.

Additionally, the naming scheme was intended to invoke a sense of inspiration by referencing celestial bodies. The model Sol complements OpenAI’s existing Daybreak program, which supports organizations in enhancing cybersecurity, although it is noted that “Sol” as a voice style in ChatGPT is unrelated and may undergo a name change.

The updated GPT-5.6 system card also conveys that all three models—Sol, Terra, and Luna—are classified at a “High” risk level for both cybersecurity and biological capabilities, requiring governance caution even for the more affordable tiers when used in sensitive industries.

Below is a comparative assessment of how these models measure up in terms of pricing against other leading models currently available:

Technological advancements: enhanced reasoning and sub-agent functionality

The primary technological development in GPT-5.6 focuses on allocating additional time and structure for challenging tasks during inference.

A new max reasoning setting is introduced in GPT-5.6 Sol to tackle problems that necessitate prolonged processing time.

Furthermore, OpenAI is rolling out an ultra mode, which utilizes subagents to distribute and expedite complex tasks, breaking away from the traditional single-agent processing model.

This new strategy has shown to enhance performance in various agent-oriented tasks, as noted in the company’s launch evaluations.

Performance benchmarks highlight significant advancements over GPT-5.5

The GPT-5.6 series shows a noticeable improvement in performance over previous models, especially in complex reasoning and long-duration tasks.

In command-line automation evaluated via TerminalBench 2.1, both Sol and Terra outperform the earlier benchmarks set by GPT-5.5, with Sol utilizing the new ultra reasoning mode to attain a record score of 91.91%, while the max mode secured 88.76%. This surpasses the previous GPT-5.5 score of 83.4% and Claude Mythos 5’s 88%.

This superiority also extends to professional workflows where Sol uniquely surpassed the halfway completion mark at 50.9% in “code mode,” while Luna marginally outperformed the previous flagship model.

In the evaluation for quantitative biology and genomics, both Sol and Terra show higher accuracy than both GPT-5.5 and GPT-5.4, with Sol achieving these results using fewer tokens.

Moreover, in cybersecurity assessments evaluating vulnerability research, the new models exceed previous performance benchmarks, with Sol achieving significantly higher exploit rates as processing time increases, while consuming a smaller number of output tokens than prior models.

According to OpenAI, on ExploitBench, Sol’s performance is comparable to Mythos Preview while generating only about one-third of the necessary output tokens.

New caching mechanics and improved speed

To aid enterprises in managing the unpredictable costs associated with running agent-based tasks, the GPT-5.6 API features a revamped caching system.

Developers can now configure specific cache points, guaranteed for a minimum duration of 30 minutes.

Under this new structure, initial cache entries are charged at 1.25 times the standard input rate, while subsequent cache retrievals enjoy a 90% discount.

This means that organizations performing repeated or similar tasks will incur higher initial costs to set up the cache, followed by significantly lower costs for reusing the cached content within the predefined time frame.

This predictability offers vital financial safeguards for businesses that frequently use extensive context windows or codebases in their tasks.

Additionally, to tackle the latency issue in enterprise environments, GPT-5.6 Sol will be launched on Cerebras hardware in July, achieving speeds of up to 750 tokens per second, catering to specialized applications requiring instant responses and advanced reasoning capabilities.

Enterprise considerations: Heightened security and algorithmic challenges

For engineering, cybersecurity, and compliance teams, deploying GPT-5.6 necessitates a thorough examination of its security framework.

To secure approval for its release, OpenAI dedicated roughly 700,000 A100e GPU hours to the automated red-teaming of GPT-5.6, focusing on uncovering “universal jailbreaks” that target systemic vulnerabilities instead of single context workarounds.

OpenAI has integrated a multi-layered protection mechanism that operates in real-time to implement intentional operational barriers for enterprise security teams.

• Model-level refusals: GPT-5.6 is fine-tuned to reject requests that could lead to malicious actions, including attempts to circumvent security protocols.

• Live misuse screening: Separate cybersecurity and biology detectors analyze outputs as they are generated.

• Activation-based screening: For Sol and Terra, OpenAI implements activation classifiers to monitor the model’s internal signals during inference. Should these indicators reveal a risk, output generation can be paused for re-evaluation. Luna does not receive this activation layer but is still monitored.

• Reasoning review pauses: When potential risks arise, content generation might cease while a more extensive system reviews the context. If deemed inappropriate, the response is blocked before reaching the user.

Given that legitimate security work often utilizes the same building blocks as offensive exploits, OpenAI acknowledges that these classifiers may lead to false positives.

Performance metrics indicate a 94.8% recall rate in biological assessments and an 81.6% recall rate in cybersecurity evaluations. These metrics provide enterprises with a rare, quantifiable insight into the safeguards, but also highlight that the system is not infallible and may occasionally overlook risks or misidentify legitimate activities.

Persistent flags can prompt automated reviews of historical interactions to determine whether users are engaging in malicious conduct or standard security protocols. OpenAI is currently exploring long-term compliance measures, including safety overrides managed by customers and privacy-preserving detection strategies to protect enterprise data from manual scrutiny.

Crucially, OpenAI emphasizes that during testing, Sol continues to be optimized for defensive applications rather than offensive strategies. In trials associated with the Chromium and Firefox codebases, it successfully isolated bugs but failed to autonomously create a functioning exploit, keeping it below the organization’s “Cyber Critical” threat level.

Nonetheless, all three GPT-5.6 models exceeded the “High” cyber threshold in controlled testing scenarios, with Sol achieving 96.7%, Terra 91.84%, and Luna 85.19%.

This distinction is significant for security-focused enterprises: OpenAI aims to position GPT-5.6 as a tool capable of enhancing vulnerability research and exploit analysis but not as a system designed for executing complex attack campaigns without human oversight under the current test conditions.

The geopolitical context of the gradual rollout

The phased rollout of the GPT-5.6 series underscores the increasing intricacies between cutting-edge AI laboratories and national security protocols.

The decision to limit early access to select vetted partners, whose information is shared with the U.S. government, is rooted in the ongoing collaboration regarding the evolving cyber Executive Order framework. OpenAI has publicly criticized this level of governmental oversight in its official communications, asserting:

“We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

This situation reveals the precarious position faced by modern tech firms. While they can leverage unprecedented efficiency and robust defense capabilities from resources like ExploitGym and ExploitBench, they must also grapple with the understanding that access to these advanced tools is often contingent upon diplomatic and regulatory approvals.