Gemini 3.7 Flash prices double on 1 January 2027

Google released Gemini 3.7 Flash three weeks after 3.6 Flash, lifting FrontierCode 1.1 Main from 34.4% to 43.6% and DeepSWE v1.1 from 49.0% to 65.3%, at $0.75 per million input tokens and $3.75 per million output tokens. That rate is introductory: it doubles on 1 January 2027 (Google). The number worth keeping from that release is not the benchmark. It is the date.

An introductory rate is a price rise with a schedule

The familiar pattern ran one way: each generation arrived cheaper than the one it replaced, and the small fast tier got cheaper fastest. A published doubling date breaks that. It tells anyone building on 3.7 Flash exactly what their unit economics look like in five months, and it prices today's rate as a launch promotion rather than as the new floor. Google is not hiding any of this, which is the point. The increase ships with the model instead of turning up in a bill.

DeepSeek raised the floor, and the cache with it

The same week, DeepSeek took V4-Pro out of preview, open-sourced its coding agent as Harness, and raised API prices to $0.66 per million input tokens and $1.98 output off-peak, double that at peak (The Decoder). The line inside that change matters more than the headline rate: a cache hit went from $0.003625 to $0.022 off-peak, roughly six times more expensive. Cache hits are the repeated part of an agent's bill, the reason a long context or a loop that re-reads the same files stays affordable. Raising the discounted read hits a coding agent harder than raising the input rate does, and it lands on precisely the workload DeepSeek has just given away.

Speed became a separate purchase

Cerebras is serving OpenAI's GPT-5.6 Sol at up to 750 output tokens per second on a new Ultrafast tier that OpenAI says runs up to 14 times faster than Standard processing, open as a limited API preview to customers including Jane Street, Podium, Basis and Rogo (Unite.AI). A tier shaped like that does the same work as the doubling date, from the other end: it takes latency out of the base price and sells it back as something a customer chooses. Read the preview list. A proprietary trading firm at the front of it is a customer selected for willingness to pay for milliseconds, not for volume.

Distribution is being sold the same way. IBM is embedding GPT-5.6, Codex and ChatGPT Work into IBM Consulting Advantage and standing up a dedicated OpenAI Practice with thousands of certified consultants and engineers, joining OpenAI's Elite partner tier (IBM). Frontier capability reaching enterprises through a consultancy is not a channel built around a falling per-token price.

The layer above is repricing the other way

In the same day's record, Bending Spoons agreed to buy Airtable all cash at a $1.29 billion enterprise value, an implied equity value of roughly $2.25 billion once Airtable's cash is added, and a sharp discount to its last private mark (MarketScale). The ledger holds the size of that discount: Airtable was marked at $11 billion in 2021 (Inc.), and is selling at roughly 2.7 times its $480 million of ARR, an 80% cut to that mark (Reuters).

So the two layers moved in opposite directions inside one week. Model providers put an increase on the calendar, raised the cached read, and began selling speed and expert support as separate goods. A software company used by more than 500,000 organisations, 80% of the Fortune 100 among them (Bending Spoons), went for a tenth of its peak mark. None of that settles where the compute costs actually sit, and none of these releases published one. It does say where the pricing power is being exercised, and this week it was not at the application layer.

What the ledger holds

The record behind this note: OpenAI, Airtable and Bending Spoons carry their valuations and funding history in one place, and the day-by-day feed is in the ledger. Every figure above links to the source it came from. Nothing here is estimated.