July 9 may be the biggest single day in AI model history. OpenAI shipped the GPT-5.6 series to everyone, and SpaceX dropped Grok 4.5 the same afternoon. For builders, GPT-5.6 is the one to read closely, because it is three models, not one, and the split matters.
Sol, Terra, Luna
Sol is the flagship for complex reasoning, coding, and long horizon agent work. Terra is the balanced pick, competitive with GPT-5.5 at half the cost. Luna is fast and cheap for high volume. Per million tokens it runs 5 and 30 for Sol, 2.50 and 15 for Terra, 1 and 6 for Luna. Most production traffic does not need the flagship. Default to Terra and reach for Sol only where the task earns it.
The features that change your code
Two additions matter more than the benchmarks. The new ultra setting runs four agents in parallel by default, trading tokens for stronger, faster results on hard tasks. And prompt caching is finally predictable, with explicit cache breakpoints and a 30 minute minimum cache life. If you run agent loops, that caching change alone can move your bill. Wire up Terra as your default, put cache breakpoints around your stable system prompts, and A/B ultra only on the tasks that keep failing.