Gemini 3.7 Flash is live at half of 3.6's price — we missed it, and OpenRouter's price is half of Google's again

Google shipped Gemini 3.7 Flash today, three weeks after 3.6 Flash, with no leak preceding it. We did not catch it, and the reason is worth publishing alongside the model.

What shipped. Tulsee Doshi's post calls it Google's most intelligent workhorse model yet for coding and agents, and frames the three-week turnaround as a result of developer feedback and algorithmic work. It is available in the API, AI Studio and Antigravity, and Gemini Spark — which reaches Google AI Pro and Ultra subscribers in more than 160 countries — moves onto it.

Google's own numbers, all measured against 3.6 Flash: FrontierCode 1.1 Main 43.6% against 34.4%, DeepSWE v1.1 65.3% against 49.0%, GDP.pdf 34.0% against 22.0%, AutomationBench 30.4% against 17.0%, and WebDev Arena Elo of 1588 against 1538. Those are first-party comparisons against the vendor's own previous model, which is the easiest kind of benchmark to win; the WebDev Arena figure is the only one with an outside scoreboard behind it.

The price, and a two-times discrepancy you should know about before you budget. Google's post states an introductory price of $0.75 per million input tokens and $3.75 per million output, available through the end of the year, which is half of 3.6 Flash's $1.50/$7.50. Two things follow. First, "introductory" and "through the end of the year" together mean this is a promotional rate with a stated expiry — plan for the possibility of a 2x increase in January rather than treating today's number as the price.

Second, and more immediately useful: OpenRouter lists Gemini 3.7 Flash at $0.375/$1.875 — exactly half what Google published today. The same halving shows up on 3.6 Flash, which OpenRouter carries at $0.75/$3.75 against Google's $1.50/$7.50. We checked this four ways. models.dev's google, google-vertex and kilo provider entries all agree with Google's post; only its openrouter entry disagrees, and that entry mirrors OpenRouter's own API, which is where we read the number directly. So it is one source out of step on two models in the same direction by the same factor, not a typo on a single row.

We are not calling either number wrong. Resellers set their own margins, and a promotional rate below first-party list is possible. But anyone sizing a budget off an aggregator right now is working from a figure that is 2x away from the vendor's published one, and that is worth knowing whichever way it resolves.

Now the part about us. When Google listed google/gemini-3.7-flash at 17:03 UTC, our registry had no entry that matched it — not armed, not live, nothing. We checked after the fact: zero of 52 armed entries and zero of every entry including live ones. No alert could have fired, and the id would not even have resolved to a model on the board.

This is not the failure we wrote about on July 21. That one was a catalog-speed problem — Google shipping first-party while the aggregators lagged, which we said we would fix with a first-party sweep and have not. Today the aggregator was fast; OpenRouter had the model within minutes of the announcement. The failure was that we had nothing to detect it with.

The root cause is a habit rather than a bug. Every Gemini entry on our board got armed because somebody leaked a name first — 3.6 Flash off an LM Arena checkpoint, 3.5 Pro off months of reporting, Omni Pro off an August leak. 3.7 arrived on cadence with nothing to leak, and a watchlist built from leaks has no row for that. Google has now shipped Flash revisions on May 19, July 21 and August 13, roughly every three weeks. The next one is predictable even though its name is not, so we have armed the version rather than the announcement: an entry that fires on a 3.8 or 3.9 id whenever it appears, sitting alongside the separate Gemini 4 hunt.

Two more things fell out of the audit we ran afterwards, and both were worse than the miss. Our alert for Mythos 5 was still armed — against a model that shipped on June 9, was export-controlled by the June 12 Commerce directive, and had those controls withdrawn on June 30. The record proving it is on our own sister site's timeline, and Axios stated it again in a piece we quoted six days ago. Three sources agreed and the registry still said "watch", which means the next sweep that saw Azure's listing would have announced a June launch as breaking news. And the entry hunting xAI's next flagship still carried 4.6 in its pattern after Grok 4.6 shipped on August 12 — a hunt that had succeeded and was one sweep from firing about its own catch.

Both are now live entries with dates, and the armed set is clean at 52 against 6,708 catalog ids. The lesson we are taking is narrower than "check more often": an alert that fires is only half-finished, because the registry still believes the thing is unreleased until someone tells it otherwise.