What happened
OpenAI released GPT-5.6 on July 9 and, in the same event, changed what ChatGPT is. The model now ships in three named tiers — Sol, the flagship; Terra, a balanced everyday model; and Luna, the fastest and cheapest — live across ChatGPT, the Codex coding tool, and the developer API.
Alongside the models, OpenAI shipped "ChatGPT Work," an agent that carries out multi-step tasks on its own, folded Codex into the desktop app, and added a service that hosts websites for paying users.
The assistant that answered your questions last year now runs the work itself — and the cheapest version of it is built to handle most of the job.
The detail almost everyone will miss
The benchmark scores will dominate the coverage. The number that should move an operator is the price.
OpenAI lists Terra at half the cost of its flagship and Luna at a fifth — $1 in / $6 out per million tokens for Luna against Sol's $5 / $30 — and claims, on its own benchmarks, that both smaller models beat a leading competitor at roughly one-sixteenth the cost.
OpenAI didn't just launch a smarter model; it built a permanent menu where the cheap tier is designed to do the expensive work — and priced it to keep customers from leaving.
The rest of the release points the same direction. A new "ultra" setting coordinates four agents working in parallel; ChatGPT Work runs on web, mobile, and desktop; Codex and a site-hosting feature now live inside the same app.
That is not a chatbox with new buttons — it's an operating surface, and the tool is now designed to finish work rather than describe it.
Why this matters if you run a business
If you shelved AI a year ago because it was too expensive at real volume or too shallow to trust with multi-step work, both objections just weakened in one release.
By OpenAI's account the tool now takes messy inputs from the places your work already lives — documents, Slack, email, shared drives — and returns finished artifacts instead of answers you still have to assemble.
The gap between "AI can chat" and "AI can do the task" is what kept most businesses stuck in pilot mode — and closing that gap is exactly what this release is built to do.
One caution: these are OpenAI's own numbers on OpenAI's own benchmarks. Independent results and your real workloads will differ, so treat the launch as a reason to re-test, not a promise.
What to do about it
You don't need to rebuild anything this week. You need to re-check three assumptions:
- Re-price your AI math. If you costed AI into a decision months ago, the per-task price of capable work has dropped since. Re-run the number before you conclude it's still too expensive.
- Default to the cheap tier, escalate on purpose. Start routine work on the cheapest model that clears your quality bar and reserve the flagship for the genuinely hard tenth. Paying top-tier prices by default is now the expensive mistake.
- Hand off one multi-step job. The agent earns its keep on tasks with several steps — reconcile these records, draft and format this report, pull this research together. Pick one, watch it run end to end, and judge it on the finished output.
The winners over the next year won't be the companies on the most advanced model — they'll be the ones who noticed the assistant became an operator and gave it real work to own. The capability is on the shelf now. Someone has to decide to use it.