Model pricing · 2026-11-12 · 321AI engineering notes · get a key
Quick answer: Gemini 3.x pairs strong multimodal defaults with very large context windows, and long context changes the cost picture: input volume explodes while output stays small. This guide does the long-context arithmetic with live floor prices, so a 1M-token context stops being a surprise line item.
Short chat traffic is roughly balanced between input and output. Long-context traffic is not: a retrieval-augmented answer might carry 400K tokens of retrieved material and return 4K tokens — a 100:1 input-to-output ratio. Since billing is linear per 1M tokens, the input line becomes nearly the whole bill. That is not a reason to avoid long context; it is a reason to price it honestly before you ship.
The flip side is that long context is also where caching pays the most: a pinned 100K-token reference billed as fresh input on every call is the single most expensive object in most stacks, and the same object billed as a cache read is a fraction of that. Long context punishes the unprepared and rewards the structured — the rest of this page is about being the second kind.
Gemini models accept images, audio and video alongside text, and those inputs are tokenized into the same context window — a scanned page costs tokens much like a text page does. Long-context multimodal jobs therefore multiply the input line twice: more modalities, more tokens per modality. The catalog entry for each model lists what it accepts and how it bills; treat multimodal volume as input volume in every formula on this page.
| Model | Market floor price | Channels |
|---|---|---|
| gemini-3.8-flash | $0.01296 in · $0.0648 out / 1M tok | 326 |
Market floor across competing suppliers, rendered live at page build (data date 2026-11-20); per-supplier quotes on the marketplace. Machine-readable: GET https://321ai.cc/api/public/model-prices.
Rendered from the live catalog at page build — floor across competing suppliers, with supplier counts. The Gemini family page carries the per-supplier detail.
Three long-context shapes, computed from the live floor at page build. Volumes are illustrative — substitute your own retrieval sizes:
| Workload | Model (live floor) | Volume | Estimated cost |
|---|---|---|---|
| RAG with heavy retrieval | gemini-3.8-flash | 0.4M in / 0.004M out | $0.0054 |
| Whole-repo analysis | gemini-3.8-flash | 1M in / 0.016M out | $0.0140 |
| Chat with long history | gemini-3.8-flash | 0.2M in / 0.008M out | $0.0031 |
Computed at page build from live market floor prices (data date 2026-11-20): cost = input tokens × input price + output tokens × output price. Platform fee is 0 and failures are never billed, so this is the whole bill. Volumes are illustrative — re-run with your own from the usage logs.
Note how little the output line contributes — at these ratios, cost control is entirely about what you send, not what you get back.
| Lever | Effect on the bill | Cost to implement |
|---|---|---|
| Cache stable prefixes | Repeated documents bill as cache reads | Prompt reorder — minutes |
| Tighten retrieval | Smaller input per call | Retrieval tuning — hours |
| Summarize rolling history | Caps per-call input growth | Pipeline change — hours |
| Route to per-token lanes | Floor price applies to every token | None — default routing |
Every number above re-derives from GET https://321ai.cc/api/public/model-prices — free, CORS-open, with updated_at. The calculator page shows the same formula applied to typical workloads, and the pricing explainer at /en/pricing covers billing semantics in depth. Failures are never billed and the platform fee is 0, so the floor you read is the bill you get.
A note on comparisons: Gemini models are often compared on context window length alone, but two models with the same window can land far apart on cost at 1M-token volumes. The comparison pages put any two models side by side on live floor prices — check there before committing a long-context pipeline to one model. For client wiring, the setup guides cover the usual tools.
The per-token rate is what it is — the window changes your volume, not the rate. Some model families price large-context tiers differently; the live catalog entry for each model states how it bills.
Retrieve tightly, cache the stable prefix, and let the marketplace floor work for you — multiple suppliers quote each Gemini model and routing prefers the cheapest healthy lane.
Text models bill per token; some image/video models bill per call. The catalog labels each model with its billing type and the current floor.
They are rendered from the marketplace cache at page build and re-derive from the public price API. For the second-by-second board, check the live marketplace.
Live prices: /en/models/ · billing: /en/pricing · setup: /en/guides/ · FAQ