Cadence is the new benchmark
Three weeks between frontier releases. Independent evaluation takes longer than that. The published numbers now describe models that no longer exist.
Gemini 3.7 Flash shipped three weeks after 3.6 Flash, moving DeepSWE v1.1 from 49.0% to 65.3% across that gap. The size of the jump is the headline. The interval is the story.
The arithmetic of evaluation
Building a harness, running it, checking for contamination and writing it up does not happen in three weeks. Neither does a security review, a procurement cycle, or a migration. Every one of those processes was designed for an annual release cadence and is being run against a monthly one.
Independent numbers now describe a model one or two versions behind whatever is being sold.
Which leaves buyers choosing between vendor results — fast, interested — and independent results — slow, honest. Neither is the thing a purchasing decision needs, which is current and disinterested.
What a board position is worth for three weeks
The same model tops Harvey LAB-AA at 90.7% and GDM-MRCR long context at 97.0%. Those are the two workloads that justified paying frontier prices — legal analysis and holding a long document without losing the thread.
A cheap-tier model taking both should be a significant event. It is hard to treat as one when the position may not survive the month, and when the price it was won at reverts on 1 January.
Why it accelerates from here
Research suggesting heavy RL post-training infrastructure is disproportionate to the problem points the same way. If the expensive part of shipping a new model gets cheaper, the gap between releases shrinks again.
What this asks of you
Stop buying on leaderboards. Build a small evaluation on your own workload that you can re-run in an afternoon, and treat vendor benchmarks as marketing with a factual basis. The team that can re-measure in a day has a real advantage over the team waiting for a third party to publish.
When the target moves monthly, the durable capability is not knowing which model is best. It is being able to find out quickly.
DataCamp — Gemini 3.7 Flash: Features, Benchmarks & Pricing → · DataNorth — Google releases Gemini 3.7 Flash for coding and agents →