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

OpenAI releases GPT-5.6 series: Sol, Terra, and Luna

Summary

OpenAI has launched the GPT-5.6 series, including Sol, Terra, and Luna, in ChatGPT, Codex, and APIs. Among them, GPT-5.6 Sol is equipped with a four-agent reasoning system, new programming tools and a high-computing power mode for complex tasks. Elon Musk, meanwhile, confirmed that SpaceXAI will release Grok 4.5 on Thursday, intensifying competition for advanced AI models.

New three-gear model lineup

According to OpenAI, the GPT-5.6 series adopts a new three-gear model architecture, named Sol, Terra and Luna respectively, replacing the previous naming method. The new system is based on capability stratification, separating model generations from intelligence levels, speed and pricing. The company said the series has been expanded to consumers, enterprises and developers and is expected to be deployed globally within 24 hours.

As the flagship model, GPT-5.6 Sol is designed for programming, scientific research, network security and complex knowledge work. OpenAI claims that compared with previous cutting-edge models, Sol completes more successful tasks while consuming fewer tokens and reducing estimated costs. The model can also check intermediate results, coordinate external tools, and correct its own work before returning to a final answer.

Four-agent inference extension advanced workflow

For high-load users, OpenAI has added two new high-power operating modes. Max mode gives GPT-5.6 more time to reason and verify answers;Ultra mode further coordinates the simultaneous work of multiple AI agents. OpenAI describes Ultra\'s default configuration for four agents to independently process different tasks, and finally integrate the results into a single output. The company says this approach is suitable for complex projects that require parallel reasoning rather than a single sequential workflow.

Developers can also obtain new workflow tools through Programming Tool Calling. OpenAI said this feature allows GPT-5.6 to write and execute lightweight programs in memory, filter intermediate information, and decide next actions without having to send every step of processing back to the language model.

Programming performance becomes a release focus

OpenAI claims that GPT-5.6Sol received the highest score in the Artificial Analysis Coding Agent Index, while also improving on Terminal Bench 2.1 and DeepSWE (benchmarks that measure command-line execution and long-term software engineering tasks). The company added that Terra surpassed Anthropic\'s Claude Fable 5 in some programming agent benchmarks, while Luna achieved better performance than Claude Opus 4.8 with lower estimated operating costs.

Enterprise functionality and security remain at the core

OpenAI said GPT-5.6 can integrate commercial software such as documents, spreadsheets, presentations, Slack, Notion, Microsoft 365 and Google Drive. When users provide templates or reference documents, Sol generates higher-quality editable presentations, financial models, spreadsheets, and written documents. In terms of security, the company released updated security information. According to the system card, GPT-5.6 Sol is classified as \"high-capability\" in the field of network security. Although Terra and Luna\'s overall performance is lower than Sol, they also reach the \"high-capability\" threshold. OpenAI said tests showed that GPT-5.6 was stronger in cybersecurity and biology than earlier systems, but neither met the company\'s \"critical\" thresholds. To manage high-risk capabilities, the company combines model training, real-time security inspections, monitoring, account-level execution and access control, while continuing to support legitimate defense efforts such as security code review, vulnerability verification, threat modeling, and software patches.

API pricing and intensified competition

In terms of API pricing, Sol charges US$5 per million of input tokens and US$30 per million of output tokens;Terra charges US$2.5 and US$15 respectively;Luna charges US$1 and US$6. OpenAI has also introduced more predictable reminder caching capabilities, including explicit cache breakpoints and a minimum cache lifetime of 30 minutes. At the same time, competition for advanced AI models continues to heat up. Elon Musk confirmed that SpaceXAI will publicly release Grok 4.5 on Thursday after completing beta testing. Musk describes it as an Opus class system that is faster, more token efficient and less costly, and debuted at the same time as OpenAI\'s GPT-5.6 release.