GPT-5.6 Went GA So I Put Its Subagents to Work

The preview period is over. As of July 9, GPT-5.6 is generally available across ChatGPT, Codex, and the API, ending that strange stretch where the strongest model in the world required government approval to touch.

The Lineup, Quick

• Sol : the flagship ($5 in / $30 out)
• Terra : everyday work at GPT-5.5 quality, half the cost ($2.50 / $15)
• Luna : the fast cheap tier ($1 / $6)

The interesting bit for me was ultra mode, which coordinates multiple agents across parallel workstreams. Sol also launched on Cerebras hardware at up to 750 tokens per second, which sounds like a spec sheet flex until you watch a planning document stream out faster than you can read it.

The Messiest Job on My Whiteboard

That's what I gave it: migrating the config layer of my TradingView alerts project from scattered JSON files into typed modules with validation. Historically this is the kind of chore I procrastinate on for months because it touches everything.

Ultra mode split it into a mapper agent cataloguing every config key, a writer generating the typed schemas, and what looked like a reviewer pass checking cross references. Total wall clock: around forty minutes.

It caught two dead settings I'd forgotten existed since March.

Pricing Snapshot While You Decide

• Explicit cache breakpoints with a guaranteed 30 minute minimum life
• Cache writes billed at 1.25x input rate, reads still 90% off
• Sol Ultrafast mode previewed for July at up to 14x standard speed

Not Flawless Though

One agent confidently renamed a key that a scheduled job still depended on, and I only noticed because my own tests yelled at me. Parallel enthusiasm apparently includes renaming things that were fine.

• Give subagents read-heavy discovery work
• Keep destructive renames behind human review
• Snapshot your repo before multi-agent runs, always
• Budget attention: parallel agents burn tokens like parallel teenagers burn money

Lesson logged: give subagents read heavy discovery work and keep destructive renames behind human review.

Bonus: GPT-Live

Same week OpenAI debuted GPT-Live, a voice model making ChatGPT Voice noticeably less robotic. More on that experiment in a few days, because voice deserves its own post.

Cheers, Yassen