Seed 1.6 Flash Guide: Cheap Multimodal Input, Short Shelf Life

Cheap models are usually sold on output price. Seed 1.6 Flash is more interesting on the input side: it takes images, video and text into a 256K window at a per-token rate low enough that feeding it large visual inputs stops being a budgeting exercise. That shape suits a particular kind of job, and it is the main reason to look at it at all.

The catch is time. It is a 2025 model, its successor family has shipped, and its OpenRouter endpoint has a listed end date. This guide covers where it fits now and how to leave it cleanly.

Where Flash sits in the 1.6 family

Seed 1.6 arrived in mid-2025 as ByteDance's general model with adaptive thinking, and Flash is the speed tier of the same generation. OpenRouter's listing describes it as an ultra-fast multimodal deep thinking model with text and visual understanding. Treat "ultra-fast" as a description of intent rather than a number: the speed you see depends on the provider serving it, so measure time to first token and tokens per second on your own prompts.

ByteDance's technical post on Seed 1.6 describes the 230B-total, 23B-active MoE base, but does not cover Flash separately. No parameter count for Flash has been published that we could verify, so do not assume it shares the base model's size.

OpenRouter lists it with a 262,144-token context, a 32,768-token output cap, image, video and text input, and support for tool calling, JSON mode, structured outputs and a reasoning toggle.

The price shape

At the time of writing OpenRouter lists Seed 1.6 Flash at $0.075 per million input tokens and $0.30 per million output. Prompts above 128K tokens are charged at $0.10 input and $0.80 output.

Two things follow. First, output costs four times input, so the model is cheapest on work that reads a lot and writes a little: captioning, tagging, extraction from screenshots, sorting video frames. Second, the long-prompt tier more than doubles the output rate. If you fill the window with a long video and ask for a long summary, the bill looks very different from the headline rate. Input vs output token pricing walks through why that asymmetry should shape what you route where.

A rough worked example: a pipeline that sends 50M input tokens of images and receives 5M output tokens of labels each day costs about $3.75 plus $1.50, so around $5.25 a day at the short-prompt rate. The same volume on Seed 1.6 would be $12.50 plus $10.00.

Flash or full Seed 1.6

On OpenRouter, Seed 1.6 costs about 3.3 times more for input and about 6.7 times more for output. The gap is largest exactly where reasoning lands, which is output.

Pick Flash when the task is perception-heavy and the answer is short: what is in this image, which frame shows the error, extract these five fields. Pick full Seed 1.6 when the answer requires multi-step reasoning, or when you were relying on its adaptive thinking to decide for you. ByteDance did not publish a direct Flash-versus-1.6 benchmark comparison that we could find, so the only reliable way to place the line is a side-by-side run on your own inputs.

Thinking on a cheap model is not free

Flash accepts the reasoning parameter, and turning thinking on can multiply the output tokens per request. On a model whose output is four times the input price, that quietly erodes the reason you chose it.

Default to thinking off for extraction and classification, and turn it on per route only where you have measured a quality gain. The token cost of reasoning models explains how to measure the difference before it shows up on an invoice.

Planning around the retirement

OpenRouter lists an expiration date of 11 November 2026 for this endpoint, the same date as Seed 1.6. OpenRouter's canonical slug points at a June 2025 build. Other providers may carry different dated builds under their own identifiers, so check which one yours serves and whether it shares the same end date.

The nearest replacement is Seed 2.0 Mini, listed at $0.10 input and $0.40 output, with the same 256K window, video input, and a four-level reasoning effort control instead of an on/off switch. That is about a third more per token, for a newer generation that ByteDance tuned specifically for high-concurrency, cost-sensitive work.

Takeaway

• Keep using Flash only for input-heavy, short-answer work that is already in production and already measured.
• Do not start new pipelines on it; the endpoint has weeks, not years.
• Re-run a sample of real traffic through Seed 2.0 Mini with effort at minimal and low, and compare labels, not just costs.
• Cut over before 11 November, and keep the old route as a fallback only until then.