OpenAI has officially unveiled its GPT-5.6 generation, introducing a tiered model structure designed to compete directly with Anthropic's Mythos class. However, the launch is overshadowed by significant friction between Big Tech and Washington: the US government has imposed strict restrictions, limiting initial access to a small group of vetted partners. This follows the drastic precedent set by Anthropic, whose Fable 5 and Mythos 5 models were pulled from the market by federal order.
A New Performance Hierarchy
With version 5.6, OpenAI adopts a layered naming scheme similar to its competitors. The suite is divided into three permanent tiers tailored for different use cases and budgets:- Sol: The flagship model, engineered for the most complex tasks and advanced reasoning.
- Terra: A balanced solution that matches GPT-5.5 performance at half the cost.
- Luna: The budget option, optimized for speed and extreme cost-efficiency (with rates up to 80% lower than top tiers).
Technical Showdown: Coding and Cybersecurity
OpenAI's published data positions Sol as a leader in agentic coding. On the Terminal-Bench 2.1 benchmark, Sol scored 88.8%, while Sol Ultra reached 91.9%, edging out Claude Mythos 5 (88%) and significantly beating Gemini 3.1 Pro Preview (70.7%).The primary battleground, however, is cybersecurity. According to The Decoder, Sol matches Mythos Preview's performance on ExploitBench (testing vulnerabilities in Google's V8 JavaScript engine) but with far greater efficiency, using roughly one-third of the output tokens. Other reports indicate Sol cleared 96.7% of cyberattack benchmarks, crossing the frontier AI cybersecurity risk threshold set by US authorities.
The Clash with Washington
The technical excitement is tempered by an access management process that OpenAI describes as "unsustainable." Concerned about the model's offensive cyber capabilities, the US government demanded a staggered and controlled release. This move aligns with broader efforts by the administration to prevent frontier models from being weaponized without rigorous screening.OpenAI has been vocal about its frustration, arguing that this government-led approval process denies essential tools to developers, enterprises, and cyber defenders worldwide. Currently, access is restricted primarily to US-based entities, though OpenAI hopes to expand the preview to international partners in supported countries like the UK and Australia shortly.
