OpenRouter's "Invisible Models": Quasar Alpha and Optimus Alpha
Recently, two "invisible models" have emerged on OpenRouter, sparking considerable interest. These are models that AI labs often release on platforms for testing before formal announcement. Users can experience them firsthand, but the developer, specific name, or background information remain unknown. This provides companies an opportunity to conduct user testing and gather feedback in real-world environments.
Quasar Alpha: The First "Invisible Model"
The first of these models to appear was Quasar Alpha. The post mentioning this model had a typo in the spelling of "Quasar".
Optimus Alpha: A Possible OpenAI Creation?
Shortly after, Optimus Alpha was introduced. According to a tweet from Testing Catalog, OpenRouter announced Optimus Alpha, speculating it was from OpenAI. It reportedly features a one million token context window and seemingly supports native image output.
Testing Optimus Alpha: A Practical Example
We can test Optimus Alpha. By enabling the switches on the page, the model can access logs and data for training. I used the chat interface, asking it to translate a piece of English text.
Translation and Context Length
The model quickly translated the text. When asked if it could provide a one million token context, it denied it, stating it was a disguised model. However, the OpenRouter description suggests it can.
Coding Capabilities: Creating a Whack-a-Mole Game
Next, I requested it to generate a Whack-a-Mole game in an 8x8 grid, using P5JS for easy previewing. The model rapidly produced basically correct code.
Customization and Troubleshooting
The game's speed was too fast, and I wanted to replace the moles with custom JPG images. I posed these two requests, which the model understood well, providing modified code and instructions. The instructions mentioned replacing "MODE JPG" with the uploaded image name. It also provided instructions on how to solve coding errors that appeared when running the code on P5JS. The model can analyze errors and provide solutions.
Speculation and Performance of Optimus Alpha
Initially, it was speculated that Quasar and Optimus were from the same company. Numerous user reports indicate Optimus Alpha excels in coding and has a 1 million token context window. This allows companies to test their products in the wild.
Sam Altman's Hint: OpenAI's Involvement?
Sam Altman mentioned excitement about upcoming features. A tweet hinted at OpenAI launching new features, potentially linked to Quasar Alpha. Altman described quasars as "very bright," suggesting the model is a "bright" or smart one.
Optimus Alpha's Strengths
Optimus Alpha delivers immediate output without any introductory text, no "thinking" preamble, and is good at debugging. It is user friendly and very capable.
Further Insights on Optimus Alpha
Matt Berman noted the AI world was "disrupted" by Optimus Alpha. He pointed out that Optimus Alpha is the second "invisible model" in a week (the first being Quasar), that appeared with no announcement, and redefined the AI landscape.
Performance and Capabilities
Berman further reported that when questioned, the model self-identified as a "Chad GPT." However, evidence suggests it might be OpenAI's rumored Oath Mini. Altman even replied to a post about it, implying it's their model. The key features highlighted include:
-
A 1 million token context window
-
Extremely fast inference
-
Currently free to use
-
High ranking in the Ella leaderboard
Berman cites its coding capabilities by referencing performance relative to other models (e.g., stronger than Gemini 2.5 Pro in certain languages).
Statistics and Summary
According to Berman, Optimus Alpha handles 26 billion tokens daily, excels in coding tasks, boasts unprecedented speed, and its million-token context allows understanding entire projects. While currently free, this may not last. The biggest mystery remains: Who created it? If not OpenAI, it would be a major surprise.
The Bigger Picture
The article recommends examining other videos, including one from Google DeepMind and AI2027 forecasts. OpenAI evaluates each new model they release and that these models require serious thought.
OpenAI's Focus
OpenAI allows the GPT-4O model to show compliance through drawing. We may be able to use its 1 million context window, which is exciting.
{# 
{# 
{# 
{# 
{#