OpenAI's GPT-5.6 Models Launch Publicly This Thursday
After weeks of restricted access, OpenAI's GPT-5.6 Sol, Terra, and Luna go public on Thursday — with benchmark wins over Claude Mythos 5 and notably lower prices.
The wait is over: OpenAI’s GPT-5.6 family — Sol, Terra, and Luna — launches publicly this Thursday. The models were unveiled in late June but initially limited to select partners while the US government’s Center for AI Standards and Innovation (CAISI) ran additional safety tests. According to Axios, the Department of Commerce has now cleared the public release.
The numbers OpenAI shared are worth a look. On TerminalBench 2.1 — a benchmark that measures how well a model handles real coding work in a terminal, the command-line environment developers use — GPT-5.6 Sol Ultra scored 91.9 percent and the standard Sol 88.8 percent, edging past Anthropic’s Claude Mythos 5 at 88.0 percent. Google’s Gemini 3.1 Pro Preview trailed at 70.7 percent. On cybersecurity tasks, Sol reportedly matched Mythos 5 while using only a third of the tokens — tokens being the small chunks of text a model reads and writes, which is also what you pay for.
That last point matters more than the benchmark bragging. Sol costs $5 per million input tokens and $30 per million output tokens. Anthropic’s Fable 5 runs nearly double at $10 and $50 — and reportedly tends to use more tokens for the same job on top. If those efficiency claims hold up in practice, the real-world cost gap could be even bigger than the price list suggests.
Here’s what’s behind the unusual launch path: GPT-5.6 is the first major model release to go through a government review before reaching the public. OpenAI openly criticized the hold, arguing it kept the best tools away from developers. Anthropic’s most capable model went through a similar process recently. What doesn’t exist yet is a binding set of rules for when and how such reviews happen — so far, it’s case by case, which is exactly what frustrates the labs.
What this means for you: If you use ChatGPT, expect the new models to show up soon after Thursday — likely with noticeable gains in coding and multi-step tasks. If you build on the API, the pricing is the headline: a frontier-class model at half of Anthropic’s rate puts real pressure on the market, and price pressure usually ends up benefiting users. A fair caveat: launch-week benchmarks come from the company itself. Wait for independent testing before rearranging your toolkit.
Sources
DeepSeek Is Designing Its Own AI Chip — and Taking Outside Money for the First Time
Reuters reports the Chinese AI lab is building an inference chip to reduce its dependence on Nvidia and Huawei, while raising $7 billion at a valuation above $50 billion.
