OpenAI Unveils GPT-5.6 Preview: Sol, Terra, and Luna Enter the Stage
· 5 min read ·
OpenAI has initiated a limited preview of its GPT-5.6 model series, introducing three distinct models: Sol as the flagship, Terra as a balanced option for everyday work, and Luna as a fast and affordable alternative. The company reports that Terra delivers performance comparable to GPT-5.5 at half the cost, while Luna offers strong capability at the lowest price point in the lineup.
A New Naming System and Tiered Pricing
With GPT-5.6, OpenAI has introduced a naming convention in which the number identifies a model's generation, while Sol, Terra, and Luna represent durable capability tiers that can each advance on their own cadence. The company says this structure gives users and developers clearer choices across intelligence, speed, and cost.
Pricing is set per one million tokens across three model sizes. Sol is priced at $5 for input and $30 for output. Terra costs $2.50 for input and $15 for output. Luna is listed at $1 for input and $6 for output. GPT-5.6 also brings more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25 times the model's uncached input rate, while cache reads continue to receive a 90 percent discount on cached input.
Additionally, OpenAI announced that GPT-5.6 Sol will be available on Cerebras at speeds of up to 750 tokens per second starting in July, with initial access limited to select customers as capacity expands.
Performance Benchmarks Across Coding, Biology, and Cybersecurity
OpenAI describes GPT-5.6 Sol as its strongest model to date, sharing evaluation results that highlight improved agentic capabilities in coding, biology, and cybersecurity. The company plans to release an expanded suite of evaluation results when the model becomes broadly available.
In coding workflows, GPT-5.6 Sol reportedly sets a new state of the art on Terminal-Bench 2.1, a benchmark that tests command-line workflows requiring planning, iteration, and tool coordination. The model also demonstrates broad improvements in biology workflows, achieving stronger results than GPT-5.5 on GeneBench v1—which evaluates long-horizon genomics and quantitative-biology analyses—while using fewer tokens.
For cybersecurity, OpenAI says GPT-5.6 Sol shifts the performance-efficiency frontier for long-horizon security tasks, including vulnerability research and exploitation. On ExploitBench², the model is competitive with Mythos Preview while using only approximately one-third of the output tokens. On ExploitGym3, a benchmark created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs, all three GPT-5.6 models show strong improvements in cyber capabilities as reasoning effort increases.
GPT-5.6 also introduces a new max reasoning effort designed to give Sol more time for deep reasoning, along with an ultra mode that leverages subagents to accelerate complex work beyond the capabilities of a single agent.
Layered Safety Safeguards and Government Coordination
OpenAI states that GPT-5.6 Sol launches with its most robust safety stack to date. The company strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, spending multiple weeks identifying weaknesses, pressure-testing the system, and hardening it against real-world attacks.
As part of its ongoing engagement with the U.S. government, OpenAI previewed its plans and model capabilities ahead of the launch. At the government's request, the company is starting with a limited preview for a small group of trusted partners whose participation has been shared with the government before a broader release. OpenAI notes that it does not believe this government access process should become the long-term default, acknowledging that it keeps tools from users, developers, enterprises, cyber defenders, and global partners. The company describes the step as a short-term measure while working with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases.
OpenAI reports that GPT-5.6 Sol does not cross the Cyber Critical threshold under its Preparedness Framework. In evaluations involving Chromium and Firefox, the model identified bugs and exploitation primitives—the building blocks of an exploit—but did not autonomously produce a functional full-chain exploit under the conditions tested. The company emphasizes that benchmark thresholds cannot capture every way a model may be used or combined with other tools, which is why it is pairing increased capabilities with stronger safeguards and a phased release.
The safeguard system is layered across multiple levels. Models are trained to refuse prohibited cyber assistance, including when users attempt to disguise intent or jailbreak the system. Real-time cyber and biology misuse classifiers evaluate output as it is generated, and for higher-risk cases, generation may be paused while a larger reasoning model reviews the conversation. Flagged activity can trigger account-level review across relevant conversations and risk signals, helping distinguish persistent malicious behavior from legitimate dual-use security work.
OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red teaming focused on finding universal jailbreaks—attacks that work across many prompts or contexts. The company also worked with third-party testers for extensive human expert red teaming, which will continue during the preview period. A rapid-response process is maintained to reproduce, assess, prioritize, and remediate newly discovered jailbreaks.
Availability and Next Steps
During the preview, GPT-5.6 models are initially available through the API and Codex to a select group of trusted partners and organizations. OpenAI plans to make them more broadly available to users of ChatGPT, Codex, and the API in the coming weeks.
The company acknowledges that users may encounter safeguards that block or refuse some requests during the preview, and that certain requests may take longer due to generation pauses for additional review. Safeguards may occasionally intervene on legitimate work, particularly in dual-use areas where defensive and offensive activity can initially appear similar. OpenAI says this is part of what the preview is designed to test, and that feedback will help reduce unnecessary blocks, improve contextual interpretation, and create a smoother experience before wider release.
OpenAI is also working with enterprise customers on longer-term safety approaches, including privacy-preserving detection, customer-operated safety controls, and access calibrated to the risk profile of individual customers, users, or workloads.
As GPT-5.6 enters its preview phase, the balance between capability and safety remains a central conversation. What are your thoughts on OpenAI's tiered model approach and its coordination with the U.S. government? Share this article and join the discussion.
