SiliconFlow Review (2026): Models, Token Prices, Verdict
An independent read of siliconflow: which models are listed, what the published token prices imply, and where a per-run media API fits better.
See the models → Official site
One endpoint over many model families
The positioning is a single API for open and commercial models, and the model grid backs that up. Listed families include DeepSeek, GLM from Z.ai, Qwen, Kimi from Moonshot, LongCat, Tencent's Hy3 and Google's Gemma, each with a release date, a context window, a maximum output and input and output prices per million tokens. The practical value of an aggregator is not the models themselves, since most are available direct, but the fact that swapping between them costs a string change instead of a new billing relationship. That matters more than it sounds once you have three or four models in production and one of them starts degrading or gets deprecated. The trade is that you inherit the aggregator's uptime on top of the lab's.
Reading the published price grid
The homepage prints real numbers, which is more than most inference platforms do. At the cheap end, Gemma-4-12B-it is listed at $0.1 input and $0.3 output per million tokens, and Hy3 at $0.132 and $0.528. In the middle sit the DeepSeek Flash variants around $0.22 to $0.44 input. At the top, Kimi-K3 is listed at $2.7 input and $13.5 output, and Qwen3.8-2.4T-A95B at $2.0 and $6.0. Output pricing is where budgets actually die: Kimi-K3 charges five times its input rate, so a chatty agent loop costs far more than the input number suggests. Price your workload on expected output tokens, not on the headline input figure. The gap between the cheapest and the most expensive entry on that grid is more than an order of magnitude, which is worth remembering before you default to the newest model in the list.
Context windows and what they cost you
Most of the chat models listed carry a 1,049K total context, with maximum output between 131K and 393K depending on the family. Hy3, GLM variants and Kimi-K2.7-Code sit lower, at 262K. A million-token window is genuinely useful for document work, but it is also the fastest way to run up a bill, because every turn of a conversation re-sends the context you have accumulated. A long-running agent on a high-priced model with a full window is the worst combination available on this price grid. If you are building agents, the cheap Flash-class entries with big windows are the ones to reach for, and the expensive reasoning models should be called deliberately rather than left in a loop.
Where the homepage stops short
Use cases are spelled out clearly enough: coding, agents, retrieval, content generation, assistants and search. What is not on the front page is rate limits, concurrency, SLA terms, data retention, or whether prompts are used for anything after the request completes. The site points to separate pricing and docs pages plus a Discord for those answers. Release dates on the model cards run through September 2026, which tells you the catalogue turns over quickly, and that any comparison you write down today has a short shelf life. Before you migrate real traffic, confirm the current price and window for your specific model on the official pricing page rather than trusting a grid on a homepage or a review like this one.
The parts worth knowing
Prices published up front
Every model card carries input and output cost per million tokens, from $0.1 input on Gemma-4-12B-it to $2.7 on Kimi-K3. Rare transparency for an inference platform, and easy to compare.
Wide open-model catalogue
DeepSeek, GLM, Qwen, Kimi, LongCat, Hy3 and Gemma families all appear in the grid. Switching between them is a model string rather than a new vendor relationship.
Long context as standard
Most listed chat models advertise a 1,049K total context with maximum output between 131K and 393K. Useful for document work, dangerous for unbounded agent loops on the pricier models.
Text-first catalogue
The homepage headline mentions text, image and video inference, but every priced card shown is a token-billed chat model. Confirm media pricing separately before planning around it.
SiliconFlow compared with Synexa
| Feature | SiliconFlow | Synexa |
|---|---|---|
| Billing unit | Per million input and output tokens | Per run |
| What is priced on the public page | Chat and multimodal LLMs with token rates | Image, video and audio models including FLUX |
| Interface | One API across open and commercial LLMs | One REST endpoint plus a Python SDK |
| Context windows | Up to 1,049K total on most listed models | Not the relevant metric for per-run media generation |
| Typical workload | Coding, agents, retrieval, search, assistants | Generating images, video and audio from a prompt or an input file |
| Cost driver to watch | Output tokens, which can be five times the input rate | Number of runs and the model you pick for each |
Evaluating it without regret
- Price your real workload
Take a week of your existing traffic, split it into input and output tokens, and multiply against two or three listed models. The ranking usually changes once output is counted properly. - Test the cheap tier first
Run your prompts through a Flash-class or Gemma-class entry before paying for a reasoning model. Many production tasks never needed the expensive tier in the first place. - Check the limits page
Rate limits, concurrency and retention are not on the homepage. Read the official docs and pricing pages, and ask in the community channel if an answer is missing. - Keep media work separate
If you also generate images, video or audio, price that path on its own. A per-run API such as Synexa is a different cost model from token billing and rarely loses to it.
Questions people actually ask
What does siliconflow cost?
Pricing is per million tokens and printed on each model card. The published range on the homepage runs from $0.1 input and $0.3 output for Gemma-4-12B-it up to $2.7 input and $13.5 output for Kimi-K3, with DeepSeek Flash variants in between. Rates change with each release, so confirm on the official pricing page.
Which models are available?
The homepage grid lists DeepSeek V4 variants, GLM-5.2 and GLM-5.3, Qwen3.8-2.4T-A95B, Kimi-K3 and Kimi-K2.7-Code, LongCat-2.0, Tencent Hy3 and Google gemma-4-12B-it, each with a release date. The full catalogue sits behind the Models link, and it turns over quickly enough that any list ages fast.
How large is the context window?
Most listed chat models advertise 1,049K total context, with maximum output ranging from 131K to 393K tokens depending on family. Hy3, GLM-5.3 variants in some entries and Kimi-K2.7-Code are shown at 262K. Check the specific model card you plan to use, because this differs per release rather than per platform.
Is this the official SiliconFlow page?
No. This is an independent review written by someone who buys inference, with no affiliation to the company. Every figure quoted here was read off the public homepage on the day it was crawled. Go to the official site for accounts, current pricing, documentation and support rather than relying on this summary.
Can it generate images and video?
The headline copy mentions inference for text, image and video, and content generation is one of the named use cases. Every priced card on the homepage, however, is a token-billed chat model. If media generation is your main workload, confirm what is offered and how it is billed before you plan around it.
What is the alternative for media generation?
Synexa handles image, video and audio models including FLUX through one REST endpoint and a Python SDK, billed per run rather than per token. That billing shape is easier to forecast when your unit of work is a finished asset instead of a conversation, and it avoids reasoning about output token counts.
Paying per run beats counting tokens
When the output is an image, a video or an audio file, per-run billing is simply easier to forecast. Synexa exposes FLUX, video and audio models through one REST endpoint and a Python SDK.
See the models →