A model discount with an expiry date is a budgeting event, not a saving

OpenAI cut the price of GPT-5.6 Sol by 50% until September 3. The discount is real. The risk is that it moves your defaults, and nothing tells you when the rate goes back.

GitHub announced this week that OpenAI has reduced the price of GPT-5.6 Sol by 50% until September 3. If you run Copilot Business or Enterprise, that is a genuine discount on a real line in your bill, and you should use it.

The part worth thinking about for more than a minute is what happens on September 4.

Which model got cheaper matters

Sol is not a mid-tier model. In GitHub's own description of the GPT-5.6 family, Sol has "the highest reasoning ceiling in the family," recommended for "complex reasoning over large codebases and demanding, long-running agentic work." Terra is the balanced default. Luna is the lightweight, lowest-cost option.

So the discount lands on the most expensive model in the family, and it is aimed at the workload that consumes the most: long-running agentic work. Agent runs are loops — read, plan, edit, re-read — so they make many model calls per task rather than one. Model choice compounds hardest exactly where consumption is already highest.

Halving the rate on that combination for thirteen days is a meaningful offer. It also makes the most expensive option in your dropdown feel affordable, precisely in the surface where it is least affordable at full price.

Discounts don't move bills. Habits do.

Here is the mechanic that catches teams out, and it has nothing to do with the discount being a bad deal.

Model defaults are sticky. A developer who switches their agent default to Sol during a promotion does not switch back when the promotion ends, because nothing prompts them to. There is no banner in the editor on September 4. There is no email. The dropdown looks identical, the model performs identically, and the rate has doubled.

What the discount really buys is a two-week window in which a more expensive default gets normalised. The saving lasts thirteen days. The behaviour it establishes lasts until somebody deliberately changes it back.

The failure mode is not using the discount — it is re-baselining on it. Bar widths are illustrative; check GitHub's published rates, which change as models are added and retired.

You will not see the change quickly

This is the part that makes a rate return harder to manage than it sounds.

GitHub's AI credit ledger is a single number per user per day. It carries no breakdown by model and no breakdown by surface, and it lags roughly two to three days behind actual consumption. There is no same-day figure at all.

Put those together and the September 4 rate return does not appear as an event. It appears as a slightly larger aggregate, several days late, with nothing in the data attributing it to a model. If anyone notices, they notice in mid-September, and the most likely explanation offered is "usage went up" — which will not be true.

Worse, if you want to compare before and after, the before has to already exist. GitHub retains most metric detail for roughly 100 days, so a baseline you did not capture at the time is not something you can go back for.

Three things worth doing in the next thirteen days

Not a maturity model. Three concrete actions, all small.

Write the date down where the decision lives. Not in a calendar reminder for one person — in whatever document records your model policy. The question to answer on September 3 is "does Sol stay the agent default at full rate?", and that question needs an owner before it needs an answer.

Capture a per-model baseline now. You want consumption per model per team for the discount window and for the two weeks before it. If nothing is snapshotting that daily, it does not exist later; the ledger will not hold the detail and the aggregate will not separate the effect.

Decide the post-September-3 default in advance, in writing. Deciding "Sol for large-codebase agentic work, Terra as the general default" costs nothing today and is nearly impossible to relitigate in October when the only evidence is a bigger invoice.

The honest version

A 50% price cut on a frontier model is good news. Use it — run the long-context work you have been deferring, do the migration you have been putting off, and get real value out of thirteen days of cheap reasoning.

Just do it knowing that the rate returns on a specific date, that nothing in your tooling will announce it, and that your own data will not tell you cleanly for another few days after that. A discount you planned around is a saving. A discount that quietly moved your defaults is a price increase with a delay on it.