Latest review · 31 October 2026 · checked 12:08 UTC
Google’s old Gemini 3.5 Flash API name now routes to 3.6. That is a model-path change, not proof of a decline.
The new route puts a concrete explanation beside ABIS’s 27 October behavior-change signal. It does not tell us whether the replacement performs better or worse on your task. Separately, one API tracker has two early warnings and no confirmed model alarms.
gemini-3.5-flash are automatically routed to gemini-3.6-flash. Capture the model name and response metadata, then compare the same safe task on the current route before calling it a quality change.What moved—and what did not
| Evidence lane | What it shows | What it does not show |
|---|---|---|
| Google Gemini API docs | Last updated 30 October: the gemini-3.5-flash string is deprecated and routes to gemini-3.6-flash. | No effective timestamp, matched result, or evidence that the route change caused a quality decline. |
| AI Stupid Level tracker | At 07:47 UTC on 31 October, 0 of 25 models had a confirmed alarm. Two early watches were unconfirmed: DeepSeek V4.1 Flash coding/tool use and Gemini 3.5 Flash Lite reasoning. | Private API-side tasks do not stand in for consumer app workflows. Flash Lite is not the same label as Google’s 3.5 Flash API alias. |
| NerfBench | Latest displayed session remains 16 October; all four tracked composites are within its 90–110% band. | Its score combines correctness, output tokens, and cost; it is not a recent pure answer-quality measure. |
| LiveNerf | Its latest completed 10-day Opus 5.5 window is labeled “no change.” | One pinned Claude Code subscription path and a fixed panel do not represent every Claude task or surface. |
Use the route as a testable explanation
If your API integration still requests gemini-3.5-flash, record that requested string and any response metadata that identifies the served model. Google says the old string now routes to 3.6 Flash. Compare one safe, repeatable task against the current documented model name with the same prompt, settings, and input. Keep correctness, latency, cost, and tool behavior separate. Google’s page confirms the product route; it does not say when the change began or whether the output got worse.
Reader reports are a different evidence lane
A new community post describes instruction-following and image-generation complaints and says the author may not renew Pro. It contains no matched before/after outputs or representative denominator. A reader can use the same approach locally: keep a prompt and output private, repeat a safe task under fixed conditions, and compare the actual work rather than the headline.
Read the full 31 October review · See current signals · Check the method