OpenAI announced today that it is previewing its GPT‑5.6 series with a small cohort of trusted partners. The effort follows a request from the U.S. government, which received an early look at the models before the broader rollout. OpenAI framed the limited release as a "short‑term step" designed to keep the company on track for a public launch in the near future.
Three versions make up the GPT‑5.6 lineup. Sol, billed as the strongest model yet, receives extra reasoning time and is positioned as the best choice for cybersecurity tasks. Terra, aimed at general‑purpose applications, delivers performance comparable to GPT‑5.5 while costing half as much. Luna rounds out the suite as the most affordable option, targeting developers who need large‑scale output on a tight budget.
Pricing reflects the tiered approach. Sol costs $5 per million input tokens and $30 per million output tokens, a steep drop from the $10/$50 rates of the previous high‑end model. Terra is priced at $2.50 for input and $15 for output, while Luna sits at $1 for input and $6 for output. OpenAI says the new rates make advanced AI more accessible without sacrificing capability.
Safety upgrades accompany the launch. The company reports that all GPT‑5.6 variants now refuse "prohibited cyber assistance," a move aimed at preventing jailbreak attempts. OpenAI spent roughly 700,000 GPU hours probing the models for universal jailbreaks and built a rapid‑response system to address any newly discovered exploits. Sol, in particular, includes reinforced protections for high‑risk activities and sensitive requests.
The preview also marks the first time the U.S. administration has examined the latest OpenAI model ahead of public release. While President Trump recently signed an AI cybersecurity order urging companies to submit powerful models for voluntary review, OpenAI noted that it does not intend for such government access to become a permanent default. The company hopes the limited partnership approach will allow it to refine the models quickly and roll them out more widely in the weeks ahead.
Industry observers note that OpenAI is not alone in granting early government access. According to a New York Times report, Anthropic, Google, xAI and Microsoft have also shared their latest models with officials, while Meta has yet to comply. Anthropic recently lifted a temporary block on its Mythos 5 and Fable 5 models after receiving government permission to redeploy them to a select set of organizations.
OpenAI’s focus on jailbreak prevention appears to be a direct response to recent incidents involving other AI firms. By dedicating extensive resources to hardening GPT‑5.6, the company aims to stay ahead of adversarial attacks and maintain trust among both partners and regulators.
Looking ahead, OpenAI plans to expand the availability of GPT‑5.6 beyond the initial trusted group. The broader release will give developers, enterprises and researchers access to a model suite that balances performance, cost and safety. If the company meets its timeline, the AI community could see a new benchmark for both capability and responsible deployment within weeks.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.