The mysterious model launched just three days ago has gone viral!
The newly launched Optimus Alpha on the large model aggregation platform OpenRouter has processed 77.2 billion Tokens, averaging over 20 billion per day.
And this number is still rising, with daily Token processing now exceeding 34 billion, ranking second and topping the Trending list.
Some users tried challenging it with MC-Bench, generating Minecraft-style scenes and comparing it with 4o-mini, with the results being immediately clear:
Others systematically tested its programming level and found that Optimus Alpha performs best in the Ruby language.
Some even directly praised that Optimus Alpha must be SOTA.
While surprised by its excellent performance, Optimus Alpha's mysterious identity has also sparked speculation...
Million-Context Window, Oriented Towards Real-World Tasks
Optimus Alpha supports a million-context window, with a maximum output of 32K.
And its response speed is fast, with a median first Token latency of only 0.81 seconds and a median output speed of 24.8 Tokens per second.
The introduction also mentioned that Optimus Alpha is mainly oriented towards real-world tasks, with special emphasis on programming.
A blogger asked it to write an e-commerce website with a shopping cart function, and Optimus Alpha designed a reasonable UI interface. The shopping cart function, which many other AIs struggle with, worked normally, and everything was fine across different files.
Or writing a Snake game, not only did it work normally, but it also added clever designs like snake head color changes and body color gradients, surpassing some other AI programming tools in innovation.
Some even used it to write an OCR text recognition application that supports handwritten text.
In terms of performance, its Elo score is 1338, ranking second on the list, just behind Claude 3.7 Sonnet, leading DeepSeek-R1 and the suspected predecessor of Optimus Alpha, Quasar Alpha.
Especially in SQL database query tasks, Optimus Alpha achieved the highest average score.
The Aider ranking shows that Optimus Alpha's programming ability is close to Quasar Alpha, Grok 3, and medium o3-mini, slightly better than GPT-4.5-preview.
Besides programming, Optimus Alpha also performs excellently in creative writing, ranking fourth in Elo score, following DeepSeek-V3.
Is the Mysterious Model from OpenAI?
The simplest and most direct way to investigate is to ask the model itself.
Since the model was released to collect feedback, Optimus Alpha is currently free to use on OpenRouter, making experimentation possible.
When asked about its identity, Optimus Alpha unhesitatingly said it was ChatGPT.
When pressed for a specific version, it responded "Based on GPT-4, with knowledge cutoff in June 2024".
Additionally, some people directly associated the name Optimus with Optimus Prime from Transformers, speculating that the mysterious model comes from Musk.
But others believe this is Altman's misdirection, and believing it comes from a Musk-owned company would play right into Altman's hands.
More convincing evidence starts with the already offline Quasar Alpha, which first appeared on the 2nd of this month.
A Reddit user discovered that when trying to use Quasar Alpha for inappropriate operations, the model's refusal method was very similar to OpenAI's.
The Tokenizer bug mentioned by this user refers to an earlier discovery that Quasar Alpha exhibited the same "read and jumble" phenomenon as GPT-4o when performing Chinese-to-English translation tasks.
This bug seems to be unique to OpenAI, not occurring on Grok, Claude, or DeepSeek.
Some even conducted more complex analysis - AI researcher Sam Paech (who initiated the previous creative writing ranking) tried to establish connections between models through differences in model responses using information science methods.
Paech found that Quasar Alpha was extremely similar to OpenAI's models, specifically pointing to GPT-4.5-preview.
Later, Altman hinted at Quasar Alpha's identity in a tweet.
Finally, returning to Optimus Alpha, tests discovered that the same bug from ChatGPT and Quasar Alpha appeared again.
Paech also has new results, adding Optimus Alpha to the latest phylogenetic tree, with the model closest to it being ChatGPT-4o updated on March 27 this year.
From the timeline, Quasar Alpha was taken down the day after Optimus Alpha went online, so some believe Optimus Alpha is a replacement for Quasar Alpha.
Besides the various signs observed in experiments, testing a new model in the community through a mysterious model has become a traditional skill of OpenAI.
Combined with Altman's hints about Quasar Alpha, the probability that Optimus Alpha comes from OpenAI is still very high overall.
As for more specific details, combined with the recently leaked "GPT-4.1" from OpenAI, which is considered an upgrade to GPT-4o, along with the confirmation from Paech's latest phylogenetic tree...
What do you think the true identity of this mysterious model is?
Reference Links:
[1]https://x.com/TheMattBerman/status/1910813233008509191
[2]https://www.reddit.com/r/LocalLLaMA/comments/1jrd0a9/chinese_response_bug_in_tokenizer_suggests/
[3]https://x.com/sam_paech/status/1910346895110848553
This article is from the WeChat official account "Quantum Bit", author: Krecey, published with authorization from 36kr.
