OpenAI's GPT-5.6 Sol launches Thursday after the U.S. government lifted its release ban. The model beats Anthropic's Claude Mythos 5 on TerminalBench 2.1 at roughly half the cost.
Key facts
- GPT-5.6 Sol scores 88.8% on TerminalBench 2.1.
- Sol Ultra hits 91.9% on TerminalBench 2.1.
- Sol costs $5/$30 per million tokens; Fable 5 costs $10/$50.
- US government lifted release ban after additional testing.
- Binding model release standards still do not exist.
OpenAI's GPT-5.6 family — Sol, Terra, and Luna — ships Thursday after a delay imposed by the U.S. government. The Department of Commerce approved the public launch following additional testing by the Center for AI Standards and Innovation, Axios reported. OpenAI had criticized the hold, arguing it kept the best tools from developers and companies.
The core claim is performance-per-dollar. On TerminalBench 2.1, Sol scores 88.8%, Sol Ultra hits 91.9%, and Anthropic's Claude Mythos 5 lands at 88%. On cybersecurity tasks, Sol matches Mythos 5 while using only a third of the tokens. Sol costs $5/$30 per million input/output tokens; Anthropic's Fable 5 runs $10/$50. [According to the Decoder], that means Sol delivers comparable or better results at roughly half the cost.
The Regulatory Vacuum
Binding standards for approving such model releases, as called for in Trump's latest AI executive order, still don't exist. The government's ad-hoc delay and subsequent lift underscore a regulatory vacuum — the same vacuum OpenAI proposed filling last week by offering Washington 5% equity in exchange for clearer rules. [As previously reported], that proposal has not yet yielded legislation.
The Competitive Landscape
GPT-5.6 arrives as Microsoft, OpenAI's largest investor ($13B+), is phasing out OpenAI models in Copilot to cut costs, replacing them with its own MAI models. AI chief Mustafa Suleyman aims to "ultimately eliminate" external model costs. Meanwhile, DeepSeek and Zhipu AI are developing custom inference chips, and Anthropic launched Fable 5 just last week. [Per Wired], Anthropic's Claude Cowork now persists on tasks even after users close their laptops, pushing toward smartphone-controlled agents.
For developers, the GPT-5.6 launch offers a cheaper, benchmark-leading alternative — but the absence of binding release standards means future model launches could face similar ad-hoc delays.
Key Takeaways
- OpenAI's GPT-5.6 Sol launches Thursday after US gov't lifts ban.
- It beats Claude Mythos 5 on benchmarks at half the cost.
What to watch
Watch for Q3 OpenAI revenue disclosures to see if GPT-5.6's lower pricing accelerates API adoption or merely cannibalizes GPT-4o usage. Also track whether the US government formalizes binding approval standards before GPT-6, or if ad-hoc delays become the norm.
Source: the-decoder.com
