OpenAI launches GPT-5.6 Sol with four-agent reasoning system

OpenAI has launched GPT-5.6 across ChatGPT, Codex, and its API, introducing GPT-5.6 Sol with a four-agent reasoning system and two lower-cost models as the rollout begins worldwide.

According to OpenAI, the GPT-5.6 family introduces a new three-tier lineup consisting of Sol, Terra, and Luna, replacing its previous naming approach with capability-based tiers that separate model generation from intelligence, speed, and pricing.

The company said availability is expanding across consumer, enterprise, and developer products, with global deployment expected to reach full availability within 24 hours.

Built as the flagship model, GPT-5.6 Sol is designed for coding, scientific research, cybersecurity, and complex knowledge work. OpenAI said the model completes more successful tasks while using fewer tokens and delivering lower estimated costs than its previous frontier models. Sol can also inspect intermediate results, coordinate external tools, and revise its own work before returning a final response, according to the company.

Four-agent reasoning expands advanced workflows

For users handling more demanding workloads, OpenAI has added two higher-compute operating modes. The company said Max gives GPT-5.6 additional time to reason through problems and verify answers before producing results.

Ultra extends that process further by coordinating multiple AI agents simultaneously. OpenAI described the default Ultra configuration as four agents working independently on separate tasks before combining their findings into a single output. The company said this approach is intended for complex projects that benefit from parallel reasoning instead of a single sequential workflow.

Developers also receive new workflow tools through Programmatic Tool Calling. According to OpenAI, the feature allows GPT-5.6 to write and execute lightweight programs in memory, filter intermediate information, and determine follow-up actions without sending every processing step back through the language model.

Coding performance remains one of the launch’s central themes. OpenAI said GPT-5.6 Sol achieved the highest score on the Artificial Analysis Coding Agent Index while also improving results on Terminal Bench 2.1 and DeepSWE, two benchmarks that measure command-line execution and long-duration software engineering tasks.

The company added that Terra exceeded Anthropic’s Claude Fable 5 on selected coding agent benchmarks, while Luna delivered stronger performance than Claude Opus 4.8 at a lower estimated operating cost.

Enterprise features and safety remain central

OpenAI said GPT-5.6 is designed to work across business software including documents, spreadsheets, presentations, Slack, Notion, Microsoft 365, and Google Drive. When users provide templates or reference files, the company said Sol produces higher-quality editable presentations, financial models, spreadsheets, and written documents.

Alongside the productivity features, OpenAI released updated safety information covering the new models. According to the company’s system card, GPT-5.6 Sol is classified as High capability in cybersecurity, while Terra and Luna also reach the High capability threshold despite offering lower overall performance than Sol.

OpenAI said testing found GPT-5.6 to be more capable in cybersecurity and biology than earlier systems, although the models remain below the company’s Critical threshold in both areas. To manage higher-risk capabilities, the company said it combines model training with real-time safety checks, monitoring, account-level enforcement, and access controls while continuing to support legitimate defensive work such as secure code reviews, vulnerability validation, threat modeling, and software patching.

API pricing starts at $5 per one million input tokens and $30 per one million output tokens for Sol. Terra costs $2.50 for input and $15 for output, while Luna is priced at $1 for input and $6 for output. OpenAI also introduced more predictable prompt caching, including explicit cache breakpoints and a minimum cache lifetime of 30 minutes.

Meanwhile, competition in advanced AI models continues to intensify. Separately, Elon Musk confirmed that SpaceXAI will publicly release Grok 4.5 on Thursday after completing beta testing. Musk described the model as an Opus-class system that is faster, more token-efficient, and lower cost, placing its launch alongside OpenAI’s GPT-5.6 rollout during a week of major AI product releases.