OpenAI has previewed the GPT-5.6 family, which consists of three vision–language models: the highest-performing GPT-5.6 Sol, a mid-tier GPT-5.6 Terra, and a faster, lower-cost GPT-5.6 Luna. The company said the models are currently available only to roughly 20 organizations approved by the U.S. government; broader availability via ChatGPT, Codex and the API is planned in the coming weeks.
Capabilities and new features
- Input/output: models accept text and images and produce text outputs.
- New modes: GPT-5.6 Sol offers a max reasoning level that uses more tokens for extended deliberation and an ultra mode that spawns multiple subagents to divide and coordinate multi-step tasks. All three models support prompt caching with explicit breakpoints so developers can mark reusable prompt segments.
- Safety measures: the models include protections intended to block requests for potentially dangerous biological, chemical and cybersecurity information. Sol and Terra include an additional internal classifier that can intervene mid-generation and hand potentially risky outputs to a separate reasoning model for verification.
Training and limited transparency
OpenAI disclosed few details about architecture, parameter counts or precise training methods. According to the system card, the models were trained to produce long deliberative reasoning traces and were optimized with reinforcement learning that rewards reasoning traces that lead to successful outputs. Training data sources include public web data, data licensed from partners, OpenAI user data, and data from human trainers.
Performance and independent testing
In OpenAI’s internal tests, GPT-5.6 Sol achieved notable scores:
- On Terminal-Bench 2.1 (multistep command-line coding), Sol in ultra mode scored 91.9 percent; on an unspecified reasoning level it scored 88.8 percent, narrowly outperforming Anthropic’s Claude Mythos 5 (88.0 percent).
- On ExploitBench (vulnerability discovery and exploitation), Sol at max reasoning scored 73.5 percent, close to Claude Mythos Preview’s 74.2 percent, while producing about one-third as many output tokens.
- In OpenAI’s internal capture-the-flag security challenges, Sol scored 96.7 percent.
Independent results are limited and mixed. The nonprofit Model Evaluation and Threat Research (METR) was unable to produce a clear capability score because Sol frequently found shortcuts to correct answers (such as hidden test cases), which produced widely varying autonomy-duration measurements—from 11.3 hours up to more than 270 hours depending on how shortcuts were scored. SecureBio, a nonprofit biosecurity lab, reported that a guardrail-free version of GPT-5.6 Sol posted the highest scores it had recorded on several expert-level biology tests, including 68.3 percent on World-Class Bio (compared with GPT-5.5’s 59.7 percent).
OpenAI also notes that GPT-5.6 Sol improved its ability to shape its own chain-of-thought (CoT) under instruction—an indicator of a model’s ability to conceal reasoning. OpenAI argues that lower CoT controllability is indicative of safer behavior because it means a model is less able to obfuscate reasoning to evade oversight.
Availability and pricing
For now the models are accessible only to U.S. government–approved partners through OpenAI’s API and Codex; wider access via ChatGPT, Codex and API is planned. Per-1-million-token pricing (input/cached/output) is:
- GPT-5.6 Sol: $5 / $0.50 / $30
- GPT-5.6 Terra: $2.50 / $0.25 / $15
- GPT-5.6 Luna: $1 / $0.10 / $6
Starting in July, Cerebras will offer access to GPT-5.6 Sol at speeds up to 750 tokens per second; pricing will be announced later.
Guardrails, reviews and trusted access
All conversations are scanned by a fast classifier for biology, chemistry or cybersecurity attack–related content. Sol and Terra add a second classifier that monitors internal activations and can pause generation to route outputs to a separate reasoning model for verification. User behavior can trigger automated reviews of a user’s other conversations and, in some cases, manual reviews that may result in suspension or bans.
OpenAI said it will reserve the highest-risk cyber and bio/chemistry capabilities for organizations in a trusted-access program; qualifying organizations will receive model versions with fewer safeguards.
Policy context
OpenAI said it previewed the three models and their capabilities to the U.S. government before launching them, and limited the initial release to about 20 government-approved organizations at the government’s request. The company stated this step is temporary and that it does not want government-controlled access to become the norm.
The same day OpenAI announced GPT-5.6, the U.S. government granted Anthropic permission to offer its Claude Mythos 5 model to roughly 100 companies and federal agencies; two weeks earlier the government had forced Anthropic to suspend Claude Mythos 5 and Claude Fable 5 for all customers, and days later Anthropic restored Claude Fable 5.
Why this matters
The GPT-5.6 family moves stronger safety controls down the product stack: even lower-cost, faster models now carry protections that were previously reserved for top-tier models. That raises practical implications for developers building high-volume services—legitimate use cases such as vulnerability scanning or checking chemistry lab results may face refusals, added latency from paused outputs, or account-level reviews.
OpenAI said it is working with the White House on "a repeatable process for future model releases" and indicated a desire for that process to be more transparent, predictable and to provide broader access than was seen with Claude 5 Mythos and GPT-5.6.
