What is 2026 Gemini 3.7 Flash: AA-AnalystAgent pass^5 and 60% score
Google amplified an independent spreadsheet-agent board on 19 August. This piece separates AA’s numbers from the product post.
On 19 August 2026 a Google account amplified an independent leaderboard: a Flash-tier model, sold as the faster cheaper workhorse, sat first on a spreadsheet-and-document agent bench ahead of several systems treated as flagships.
The object is Gemini 3.7 Flash. Artificial Analysis’s AA-AnalystAgent scores pass^5: each item is run five times and counts only if all five are correct. What follows is the AA board, AA’s 13 August note, Google’s 13 August product post and OfficeChai’s recap — not a claim that Flash now leads Opus or GPT in general.
What the test is
AA writes that the bench targets quantitative questions a business or data analyst would face: 80 tasks across 14 business and scientific domains, each pinned to a folder of real spreadsheets and documents rather than a cleaned dataset. The question set and reference answers stay private aside from a few examples, to limit contamination. Results are a standalone leaderboard, not folded into the Intelligence Index.
Figures that can be checked
- 1
Ship date is 13 August; the amplification is 19 August
Google’s Introducing Gemini 3.7 Flash post dated 13 August, about three weeks after 3.6 Flash, and called it a workhorse for coding and agents. AA’s same-day article already listed AnalystAgent (high) at 60%. NewsFromGoogle on 19 August claimed the top rank and wrote 60%–90% faster than other top models, and 2.4× the closest accuracy rival.
- 2
On this board Flash sits above several flagships
AA lists Flash (high) 60.0%, then Claude Opus 5 (Adaptive Reasoning, Max Effort) 53.8% and GPT-5.5 (xhigh) 50.0%. OfficeChai also listed Claude Fable 5 (fallback) 48.8%, GPT-5.6 Sol (max) 47.5%, Grok 4.6 41.3% and Kimi K3 38.8%. That is this bench, not a general-index flip.
- 3
On the general intelligence index it is not first
AA’s 13 August note: Flash (high) scores 56 on the Intelligence Index, behind GPT-5.6 Terra (max) and Muse Spark 1.2 (xhigh) at 57, just ahead of Claude Sonnet 5 at 55. Average Time per Task is 1.7, about 40% faster than Terra (max), on the intelligence-versus-time Pareto frontier.
Google is also selling a speed-and-price story
The product post lists an introductory price through year-end: $0.75 / 1M input and $3.75 / 1M output, then $1.50 / $7.50 from 1 January 2027. AA’s model page lists a 1M context and the same input price. Price is list pricing, not a variable inside this bench.
AA reported a second agent board
The 13 August AA article says Flash (high) leads AutomationBench-AA (agents in simulated SaaS) at 62.7%, ahead of Kimi K3 (max) 53% and GPT-5.6 Sol (max) 51.2%. That is a different set. It should not be collapsed into “first everywhere.”
Items stay private, so reproduction is limited
AA says the question set and keys are held privately. What can be checked is an independent lab’s pass^5 under its protocol, not a shared folder anyone can rerun. Contamination risk is lower; outside audit is harder.
| Claim | Checkable source | How to read it |
|---|---|---|
| 3.7 Flash shipped 13 August | Google product post; AA model page | Ship date, not a 19 August launch |
| AnalystAgent pass^5: 60.0% / 53.8% / 50.0% | AA board; OfficeChai recap | First on this bench, not the general index |
| 2.4× the closest accuracy rival | NewsFromGoogle 19 August (via OfficeChai) | Company amplification; AA also cites 1.7 Time per Task |
OfficeChai wrote that Flash is not the Intelligence Index leader, but currently has the strongest published claim on “pull the same defensible number five times from messy documents.” That is commentary, not AA folding AnalystAgent into the index.
# AA-AnalystAgent (reported 2026)
80 tasks × 14 domains
pass^5 = correct on 5/5 runs
Gemini 3.7 Flash (high) 60.0%
Claude Opus 5 (max) 53.8%
GPT-5.5 (xhigh) 50.0%How to read the boundaries
- A specialist agent board ≠ the flagship is obsolete
- 3.7 Flash still sits below Pro in the naming. Google’s post also called it a workhorse, not the next frontier general model. First on AnalystAgent does not auto-win coding, cyber or chat boards.
- A high pass^5 still misses about two fifths of items
- 60% is about 32 of 80 tasks correct on all five runs. AA does not publish a single-run accuracy as the same headline. It should not be read as “60% of analyst jobs can run unattended.”
- Speed has two published framings
- Google social copy said 60%–90% faster and 2.4×. AA cited Time per Task versus Terra. Neither is a notarized stopwatch on the same 80 tasks by a third party.
Questions worth checking
Does this mean Gemini now leads Opus and GPT overall?
No. AA’s own Intelligence Index still lists Flash (high) at 56, behind Terra and Muse Spark 1.2 at 57. What can be checked is rank on AnalystAgent and AutomationBench-AA.
How is pass^5 different from ordinary accuracy?
AA writes that a task counts only if all five attempts succeed. One lucky solve does not lift the headline. The set is private, so outsiders cannot recompute it from the same folders.
When did 3.7 Flash ship, and at what price?
Google’s post says 13 August, introductory $0.75 / $3.75 per million tokens through the end of 2026, then $1.50 / $7.50 from 1 January 2027. AA’s model page also lists 13 August 2026. Regions and product doors follow the official pages.