I run a pricing check every night against the vendors’ own rate cards — OpenAI, Anthropic, Google, xAI, DeepSeek, Moonshot, Z.ai, Meta. Over the last month the check came back “no change” almost every night. That is the headline most people would write. It would also be wrong. The stable sticker prices are hiding where the actual action went: new tiers underneath the flagships, and a consumption mix that is tipping decisively toward cheap models.

Start with the one real rate-card change: GPT-6.1 Terra, launched October 26 at $1 per million input and $5 per million output. It is the first Terra-class model since GPT-5.6, and it slots half a rung below GPT-6.1 Sol’s $2/$10 — a family that already cut its own cached-input price to $0.10 when it arrived in late September. Terra’s positioning matters more than its number: OpenAI aimed it at high-volume routine work like invoice processing, and quoted roughly nine times lower cost per task than Opus 5.5 on its own benchmarks. Vendor numbers, so read them with that grain — but the direction is unmistakable. The frontier is not getting cheaper. The mid-tier is getting closer to the frontier, and the price of “close enough” keeps falling.

Anthropic held completely still. Fable 5.1 at $10/$50, Opus 5.5 at $4/$20, Sonnet 5.5 at $2/$10, Haiku 5.5 at $0.10/$0.50 — every row verified against the pricing page this week. Google’s Gemini 4 Argon went generally available on October 13, but that was an availability change, not a price change: the introductory $2/$10 rate runs through November 30, then steps up to $4/$20 on December 1. Mark that date. Kimi added K3.1 Pro at $2/$10 — the same band as Sol, Opus 5.5, and Meta’s Muse Wonder — with a Highspeed variant at $4/$20. Grok 4.7 has not moved since September 21.

The second story is who is actually being used. OpenRouter’s public rankings give token consumption per model, and this month every spot in the top ten belongs to a model whose output costs under about $5 per million tokens. The volume leader is DeepSeek’s V4.1 Flash at $0.60 off-peak output — 35.4 trillion tokens routed in a week, up 382%. GLM-5.3 Flash ($0.50) and MiniMax M3.1 Flash follow. The only premium model with meaningful volume is GPT-6 Luna at $0.50 output — and Luna is OpenAI’s own cheap tier, which is itself the point. The most expensive model on the board, Claude Opus 5.5 at $20 output, moved 8.23 trillion tokens, roughly a quarter of the leader’s volume. That is a real, expensive, paying habit — but it is one row among ten, not the whole board.

Two models crossed the trillion-token mark for the first time this month and are worth naming: Kimi K3.1 Pro, which hit the board within days of its October 23 API launch, and GPT-6.1 Terra itself, which is already routing more than seven billion tokens. New models are not waiting months to find users. They are landing with usage attached.

The nightly “no change” is now the norm, and it is misleading in a specific way. The listed prices are not moving; the ladder underneath them is being rebuilt, rung by rung, and the workload mix is draining toward the bottom of it. If your cost model still assumes the flagship is where the volume goes, it is a month out of date. The December 1 Argon step-up and the January 1 Flash reset are the next scheduled changes. The unscheduled ones are the ones to watch.