Signals · 31 October 2026 · checked 12:08 UTC
The Gemini API route changed; the two current tracker warnings are not confirmed alerts.
Dates, exact model strings, and what each score measures change the reading. No reviewed source establishes broad capability decline across popular models.
| Source | Latest observed | Limit |
|---|---|---|
| Google Gemini API documentation | Updated 30 October: gemini-3.5-flash is deprecated; API calls to that name route to gemini-3.6-flash. | Confirms an alias/serving change, not a decline, its exact effective date, or the cause of ABIS’s 27 October behavior signal. |
| AI Stupid Level | 31 October, 07:47 UTC: 0 open alarms among 25 models; two unconfirmed early watches; 66/75 per-suite detectors armed; 3 models calibrating. | DeepSeek V4.1 Flash coding/tool use and Gemini 3.5 Flash Lite reasoning are only early warnings on operator-run API tasks. |
| LiveNerf / Claude Opus 5.5 | Latest completed 10-day window: −0.015 ± 0.018 over 78 items, labeled “no change”; later window was in progress in a 29 October status. | One calibrated panel through pinned Claude Code subscription path; not every Claude product or task. |
| BridgeBench / NerfBench | Latest visible test: 16 October; four model series remain within the stated 90–110% composite band. | Old snapshot; composite mixes correctness, output tokens, and cost. First tracked session is the baseline. |
| AI Drift Detector | Product page describes weekly tracking; the visible board snapshot is 27 September. | Stale displayed date; the current model ranking did not render in the inspected page text. |
How to read the Gemini signal
ABIS’s 27 October finding labels Gemini 3.5 Flash a possible version-update signal. Google’s newly inspected API page confirms that the old gemini-3.5-flash API name now routes to 3.6 Flash, which makes a serving-path change plausible. The sources do not give an effective date or matched outputs, so this is a possible explanation, not a causal finding. The tracker’s watch is for the distinct Gemini 3.5 Flash Lite model and should not be folded into the same event.
Before you react
For a changed API response, preserve the requested name and any response metadata, then repeat one safe task with the same prompt and settings on the documented route. For a stopped request, check service status. For a poor completed answer, score task correctness and correction burden separately from route, time, price, and output length.