OpenAI Cuts Sol's Price 20% While It Leads Terminal-Bench

OpenAI cut GPT-5.6 Sol's price 20% on August 21, 2026 — input from $5 to $4 per million tokens, output from $30 to $20 — and flagged the new rate as promotional through at least November 21. It is the third cut in the GPT-5.6 line in under a month. Terra dropped 20% on July 30. Luna dropped 80% the same day and now lists at $0.20/$1.20. The frontier tier is now being priced like the mid tier was six months ago.

The timing is what makes this interesting rather than routine. Sol posts 89.5% on Terminal-Bench 2.1, edging Claude Opus 5's 89.1% on the same benchmark. So OpenAI's most expensive current model just got cheaper in the same window it took the lead on an agentic coding benchmark against the model helloai currently ranks as the strongest coder. On raw price-per-capability, that is a real argument, and it is the first one OpenAI has had in a while.

It is not, however, an admission argument. helloai's tracked six is gated on Elo — LMArena head-to-head preference — and nothing else. Sol left the table on July 17 when Kimi K3 replaced it, and Sol was the lowest-Elo entry at the time at 1484. Terminal-Bench measures something narrow and useful: can the model drive a terminal to completion. Elo measures something broad and noisy: do humans prefer the output across everything they happen to ask. A win in the first does not move the second, and the tracked set only reads the second.

There is a wrinkle worth sitting with. The current tracked floor is Grok 4.6 at 1461 — below the 1484 Sol carried when it was dropped. Read one way, that is an inconsistency: the bar moved down after Sol left, and Sol would clear today's floor. Read another way, the July decision was made against July's field, admission is not retroactive, and re-litigating past cuts every time the floor drifts turns a ranked table into a queue of appeals. Neither reading is obviously wrong, which is usually a sign the rule is doing more work than the number behind it.

What the price cut actually changes is the invoice, and that is not nothing — at $4/$20 with a Terminal-Bench lead, Sol is a defensible pick for agentic coding workloads regardless of where it sits on any directory's table. The open question is whether an Elo-gated set should have a price input at all. If cost-per-capability never touches admission, the tracked six will keep describing which model people prefer while saying nothing about which model a team should actually buy. If it does touch admission, the criterion stops being one number and starts being a judgment call. OpenAI has spent a month making that question harder to avoid.