Last updated 07-30-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
Avian
Avian sells access to open weight frontier models through an OpenAI compatible API, billed by the token. Point an OpenAI SDK client at the https://api.avian.io/v1 base URL and name a model like DeepSeek-V3.2, and the rest of the code stays as it was. Everything runs on NVIDIA B200 GPUs with speculative decoding, and Avian measures 489 tokens per second on DeepSeek V3.2.
The argument against OpenAI here is arithmetic rather than features. A million output tokens on DeepSeek V3.2 costs $0.33, against the $10.00 Avian quotes for GPT-4o, with no rate limits, no monthly minimum, and prepaid credits that never expire. What you give up is model choice, since the catalog is entirely open weight families, DeepSeek, Kimi, GLM, and MiniMax, with no GPT, Claude, or Gemini anywhere.
Developers wiring up coding agents get the clearest win, because Cursor, Claude Code, Cline, Windsurf, and Kilo Code all accept a custom OpenAI base URL. Teams with data handling rules will care that prompts and completions are processed in memory and never written to persistent storage, with only timestamps, model name, and token counts kept for billing.
One claim deserves a closer read. The homepage advertises SOC/2 approved infrastructure while the security page says Avian is still in the process of obtaining SOC 2 Type II, so that badge is work in progress rather than a finished audit.
OpenAI compatible endpoint at https://api.avian.io/v1 turns moving an existing integration into a one line base URL change
Catalog covers DeepSeek V4 Flash and Pro, DeepSeek V3.2, Kimi K2.5 and K2.6, GLM-5, GLM-5.1, and MiniMax M2.5
Rates start at $0.0945 per million input tokens on DeepSeek V4 Flash, with cached input as low as $0.012 per million on V3.2
Measured 489 output tokens per second on DeepSeek V3.2, running on NVIDIA B200 GPUs with speculative decoding and no cold start
Prepaid credits in $50, $100, $150, and $250 top ups that never expire, with no rate limits while a balance remains
Prompts and completions are processed in memory only, with timestamps, model name, and token counts the sole logged metadata
Dedicated deployments on reserved NVIDIA H200 or H100 GPUs are quoted through sales
Output tokens run about a tenth of the GPT-4o rate at $0.33 per million on DeepSeek V3.2, using Avian's own comparison
An existing OpenAI SDK integration moves across with a base URL change instead of a rewrite
Prepaid credits never expire and carry no monthly minimum, so a quiet month costs nothing
Prompt and completion content is never written to disk, which matters where sending content to a logging provider is not allowed
No rate limits while credits remain, so throughput is not gated behind a higher tier
Catalog is limited to DeepSeek, Kimi, GLM, and MiniMax, with no GPT, Claude, or Gemini models available
The homepage advertises SOC/2 approved infrastructure while the security page says SOC 2 Type II is still being obtained
Credits are sold in $50 increments up front, which is a large first commitment for a small evaluation
Speed and cost comparisons against OpenAI and Anthropic are published by Avian with no third party benchmark cited
How much does Avian cost?
Avian bills per token from a prepaid balance with no subscription. DeepSeek V4 Flash runs $0.0945 per million input tokens and $0.189 per million output, while GLM-5.1 is $1 input and $3.20 output. Credits come in $50, $100, $150, and $250 top ups and never expire.
Which models does Avian serve?
Avian serves DeepSeek V4 Flash, DeepSeek V4 Pro, DeepSeek V3.2, Kimi K2.5, Kimi K2.6, GLM-5, GLM-5.1, and MiniMax M2.5. Context windows run from 163K tokens on DeepSeek V3.2 up to 1M on the V4 models, and one API key reaches all of them.
Is the Avian API OpenAI compatible?
Yes. Avian follows the OpenAI Chat Completions format, so pointing an OpenAI SDK client at https://api.avian.io/v1 with an Avian key is the only change required. Streaming, native tool calling, and vision requests use the same request shapes.
How fast is Avian inference?
Avian measures 489 output tokens per second on DeepSeek V3.2, comparing that to 120 for GPT-4o and 90 for Claude 3.5. Models run on NVIDIA B200 GPUs with speculative decoding and stay pre-loaded, which the company reports as a 0ms cold start.
Does Avian store prompts and completions?
No. Avian states that prompt and completion content is processed in memory and never written to persistent storage. It logs only request metadata such as timestamps, the model selected, and token counts, used for billing and service monitoring.
Does Avian rate limit API requests?
Avian states there are no rate limits, and that the prepaid balance is the only cap on request volume. When credits run out, requests stop until the balance is topped up again.
Which coding tools work with Avian?
Avian names Cursor, Claude Code, Cline, Windsurf, and Kilo Code among more than 20 supported coding tools. Any client that accepts a custom OpenAI base URL can point at https://api.avian.io/v1 and authenticate with an Avian key.