Aeion AI + Google Gemini
Configure Google Gemini as a provider in Aeion AI. Covers Gemini 3.1 / 3 and the still-active 2.5 family, text + image input plus audio transcription, up to a 2M-token context window, and the cost-leader Flash family for bulk workloads.
Setup
Aeion talks to Gemini through the public Generative Language API.
Model Family — When to Use Which
Gemini 3.1 / 3 are the current flagship + fast tiers; Gemini 2.5 remains active for existing workloads.
Multimodal Input — What's Actually Wired Up
Gemini itself natively accepts text, image, audio, and video in a single request. Aeion's Google adapter currently wires up text + image input through the standard completion path, plus a separate audio-transcription operation — video and PDF as chat input aren't implemented yet.
Cost Math vs OpenAI / Anthropic
Same 50M-token-month workload across three providers, all-bulk (assuming a 3:1 input:output ratio).
Common Gotchas
- Safety filters block content — Gemini's safety filter is strict by default. Configure thresholds via the
safetySettingsprovider option; defaults block at "MEDIUM" probability for harassment, hate speech, sexual, dangerous content. Adjust to "ONLY_HIGH" or "NONE" for adult-content apps.
FAQ
Yes — Aeion's unified tool-use abstraction translates to Gemini's `functionDeclarations` field automatically.
No — Aeion's Google model catalog has moved on to the 2.0/2.5/3 generations. If you were relying on a 1.5 model string, move to `gemini-2.5-flash` or newer.
Linear — at 1.5M input tokens × $1.25/1M = $1.88 per request for Gemini 3.1/2.5 Pro. Compare to Claude's 1M-context preview at a much higher per-token rate. Gemini Pro is materially cheaper for long-context work.
`text-embedding-004` is the current default and cheapest option. Aeion's `embed()` method handles it; `gemini-embedding-2` is also available for higher-dimensional embeddings.