OpenRouter surprised the AI world by quietly releasing a new model called Quasar Alpha on April 4, 2025. This quiet launch caused a lot of talk about where it came from and what it can do.
According to OpenRouter's announcement, Quasar Alpha is a prerelease of a long-context foundation model from one of their partner labs. It has a huge 1 million token context window. This window is built mainly for coding but is also good for general tasks.
Unlike most flashy model launches, Quasar Alpha arrived quietly. OpenRouter described it as a "stealth" model available for free. It is aimed at developers who need advanced long-context coding features.
Update: Quasar Alpha confirmed to be an early version of GPT-4.1
OpenRouter has confirmed that Quasar Alpha is an early test version of GPT-4.1 from OpenAI.
We also found evidence that GPT-4.1 is closer to the second stealth model Optimus Alpha than the original Quasar Alpha. This is evident in the Long Context benchmark by Fiction.live.
GPT-4.1 is also slightly more verbose than Quasar Alpha in our own internal evaluation.
Strong Performance and Following Instructions
According to Paul Gauthier, The mysterious Quasar Alpha scored 55% on the aider polyglot coding benchmark, competitive with o3-mini-medium, the latest DeepSeek V3 and Claude 3.5 Sonnet.
Early feedback shows Quasar Alpha has strong skills, especially in following instructions. According to paradite_ on X, it follows instructions better than Claude 3.5 Sonnet and Gemini 2.5 Pro.
paradite_ also noticed that its style is very close to GPT-4o, OpenAI's current top model.
This close style makes people wonder if Quasar Alpha is from OpenAI, but under a different name.
Technical Clues Suggest OpenAI Is Behind It
Several people in the AI community looked into the technical details. One important clue is from the generation metadata seen during API use.
As ZYPX4 pointed out, Quasar Alpha generations use an upstream ID starting with "chatcmpl-". This prefix is only seen in OpenAI API calls, like ChatGPT and GPT-4 outputs.
Also, lowvram saw that Quasar Alpha's tool call ID format is just like OpenAI's style. Companies like Google and Mistral use different styles.
These clues do not officially prove anything. Still, they suggest Quasar Alpha connects directly to OpenAI's model systems. The metadata points strongly to OpenAI quietly releasing a test version under a different name.
Clustering Models Shows Quasar Alpha Is Close to OpenAI
To learn more, AI researcher Sam Paech used bioinformatics clustering tools, known as PHYLIP pars, on model outputs.
This method checks how models are related by finding small differences in their replies. PHYLIP pars looks for the simplest family tree of models, unlike normal clustering.
Sam found that Quasar Alpha is very close to OpenAI models, specifically GPT 4.5 Preview, and is quite different from other models. This makes it more likely that OpenAI is behind it.
Chinese Response Bug in Tokenizer
A Reddit user nekofneko found a translation bug in the response of Quasar Alpha using Cherry Studio via OpenRouter.
This bug was due to the OpenAI tokenizer (o200k_base) not handling Chinese characters correctly, encoding the entire phrase "给主人留下些什么吧" as a single token with ID 177431. And it was previously reported for other OpenAI models like GPT-4o.
This is further evidence that Quasar Alpha is from OpenAI, although it is possible that other labs are also using the same tokenizer, or this is a bug with the OpenRouter API when used with Cherry Studio.
Update 1: Alternative Theory for a Smaller Lab
There is a new alternative theory suggesting that Quasar Alpha is not from OpenAI, but from a smaller lab called SILX AI.
There are a few circumstantial clues to support this theory:
- The models hosted on their Hugging Face are called Quasar series.
- It is much more likely for a smaller lab to collaborate with OpenRouter compared to a large lab like OpenAI.
- A Discord user called TroyGPT who claims to be from SILX AI has been actively engaging with other users on the topic of Quasar Alpha in the OpenRouter discord server.
However, this theory is unlikely to be true due to the following reasons:
- OpenRouter was serving about 67.4B tokens a day at peak for this model, which is unlikely for a small lab to handle without a huge amount of compute resources.
- The discord user TroyGPT only joined the OpenRouter discord server on April 4, 2025, the same day Quasar Alpha was launched, which means it is unlikely that SILX AI has been working with OpenRouter to coordinate the launch.
Update 2: Further Evidence for OpenAI
X user jakobdylanc has found further evidence for OpenAI being behind Quasar Alpha.
In the chat completions API spec, the message object typically only has 2 fields: "content" and "role". However, Quasar Alpha's message object supports an additional "name" field, which is only supported by OpenAI and xAI.
Further more, the support for the "name" field is not consistent across all providers, and xAI supports any characters, while OpenAI only supports alphanumeric characters. During testing, jakobdylanc found that Quasar Alpha only supports alphanumeric characters, which is consistent with OpenAI's behavior.
Why the Stealth Release?
There are some reasons why an AI lab would stealth launch a model without major publicity.
Testing in real-world conditions without hype allows for gathering honest developer feedback, while maintaining a low profile reduces pressure to meet high expectations.
A stealth release can also encourage a more fair and unbiased comparison with other models in the market, without the distraction of marketing claims.
What Quasar Alpha Offers Developers
Quasar Alpha offers developers a rare chance to work with an impressive 1 million token context window, especially suited for coding and also general tasks.
Its reported strong instruction following, combined with extensive context handling, provides a valuable tool for those working on complex software or large documents.
Try Quasar Alpha in 16x Prompt
You can try out Quasar Alpha in 16x Prompt via OpenRouter.
Simply follow our OpenRouter setup guide and select Quasar Alpha as the model.
You can also compare its output against other models like Claude 3.5 Sonnet and Gemini 2.5 Pro.
