The cheap-model market flipped this week, and the frontier labs are the ones who flipped it

Three things landed inside about 36 hours, and they only make sense read together.

OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%. DeepSeek shipped an official V4-Flash that scores 82.7 on Terminal-Bench 2.1 at $0.14/$0.28. Artificial Analysis has Luna at 51.2 on its Intelligence Index v4.1 — the top score in the budget tier — at the lowest input price in that tier.

For two years the shape of this market was stable: frontier labs sold intelligence, Chinese open-weight labs sold price, and you picked your axis. That trade is gone. The highest-scoring cheap model is now also the cheapest cheap model, and it belongs to OpenAI.

Read the direction of causation carefully, because the press mostly got it backwards. OpenAI did not cut prices because it wanted to. V4-Flash preview was already at 40 on the AA index in April; the official version is at 50, matching Gemini 3.6 Flash. GLM-5.2 Max sits at 51. When four labs can put a ~50-scoring model in front of you and one of them charges $0.14, a $1.00 input price stops being a margin and starts being a reason to leave.

What this means if you buy inference: the budget tier is now genuinely frontier-adjacent for agent work, and the differences that remain are not captured by an index number. They are latency, throughput, uptime, rate limits, and whether the price you signed up at is the price next month. Those are exactly the surfaces nobody publishes a leaderboard for — and exactly what we watch.

One caution on the charts going around: the AA Intelligence Index is a composite, and a composite that moves 10 points on a post-training-only release is measuring the harness as much as the model. Treat 50 vs 51 as a tie.