Gemini 3.7 Flash vs 3.6 Flash — same perfect grades, 2.6x faster, 45% cheaper, and the 5 test prompts to prove it on your own work
The companion to the video. On August 13, 2026 — day one for Google's Gemini 3.7 Flash — I ran it head-to-head against the model it replaces, Gemini 3.6 Flash: 5 real jobs, identical prompts, temperature 0, machine-graded so no opinion touches the scoreboard — and the full battery run three times end-to-end, with every number reported as the median of the three runs. Both models scored PERFECTLY on every job, every run — a buggy prorated-refund billing function (both 400/400 hidden tests), a single-file ops dashboard built from a 9-requirement spec (both 12/12 automated browser checks), a real BEA GDP release PDF with 10 questions including the Strategic Petroleum Reserve trap (both 10/10), an agentic restock run with tool calls and the on-hand-minus-promised trap (both filed the exact right purchase orders), and a hard-mode cron scheduler debug with the day-of-month/day-of-week either-match rule (both 200/200 against a brute-force oracle). The difference is the meter: $0.063 vs $0.114 total (45% cheaper), 39.5s vs 1:43.1 of model time (2.6x faster), less than half the thinking tokens — and thinking bills as output. INSIDE (4 pages): (1) THE SCOREBOARD — every job, grade, time, and cost on one page. (2-3) THE 5 TEST PROMPTS, verbatim (shortened only where marked), each with a swap-in note and a grading recipe so you can rerun my exact day-one test on any model with your own work. (3) THE ROUTING RULE — if you can describe the job AND check the result, it's a workhorse-model job; open-ended or undefined, pay for a frontier model. (4) THE JAN 1 CHECKLIST — today's $0.75/M-in $3.75/M-out price is an intro price through Dec 31, 2026, and it DOUBLES to $1.50/$7.50 on Jan 1, 2027: budget at the doubled price, model string gemini-3.7-flash, knowledge cutoff ~March 2026 so paste fresh context. Pricing verified on Google's official pricing page the same day. Independent test from Hyperautomation Labs — not affiliated with Google.
Free. No spam. Unsubscribe anytime.