I don't know whose AI model this is, but it's giving away 1 million context for free—I hooked it up with Hermes. Lately, there's a pretty interesting trend in the AI world: every big model launch gets more extravagant, with parameters hyped up bigger and bigger, but what really gets people hooked are actually a few "nobody's heard of" anonymous models. On August 20th, a model called OxAlpha quietly popped up on OpenRouter.

💡 What You Will Learn

I don't know whose AI model this is, but it's giving away 1 million context for free—I hooked it up with Hermes. Lately, there's a pretty interesting trend in the AI world: every big model launch gets

📜 Table of Contents

• The Spec Sheet Is Ridiculous (And Free)
• Who's Behind It, Anyway? (Nobody Knows)
• How I Integrated It (Three Steps)
• Before You Freeload, Remember These Two Things

An AI Model From Nobody Knows Who, Giving Away 1M Context for Free — I Hooked It Up with Hermes

There's been a pretty interesting trend in the AI world lately: the big model launches are getting more extravagant, the parameter counts are getting hyped up bigger and bigger, but what's really getting people hooked are a few anonymous models that are practically "nonexistent."

On August 20th, a model called OxAlpha quietly popped up on OpenRouter. No official announcement, no model card, not even any clue who developed it — the provider field just said two words: Anonymous.

By all logic, people should be steering clear of a model with such shady origins. But what happened? Within less than a week of going live, it shot to #3 on OpenCode's usage charts, consuming 7.1 trillion tokens in a week — about 2% of the platform's total traffic. The only ones ahead of it were DeepSeek V4 Flash and Xiaomi's MiMo — both legitimate, named models from major companies.

The Spec Sheet Is Ridiculous (And Free)

My jaw dropped a little the first time I opened its parameter page:

What does a 1M context window even mean? Roughly enough to stuff an entire medium-sized codebase in one go, or a full-length novel of around 300,000 Chinese characters. Currently, models on the market with a 1M context, like DeepSeek V4 Pro, charge a couple of bucks per million tokens for combined input and output. OxAlpha just slapped a big fat 0 on it.

Even rarer is video input. You can count the models that support "watching video" on one hand, and most free models don't even handle images. Throw in forced reasoning — this model does a deep thinking pass before every single answer, making it the "thinks slow but thinks deep" type.

Who's Behind It, Anyway? (Nobody Knows)

This is the most fascinating part about OxAlpha.

It's a classic "stealth" (anonymous preview) model: some major company hangs an unreleased new model on an API platform, collects feedback from real users, and then officially announces it once they've gathered enough. This kind of play has happened 4 times in the past six months, and every single time, it was later revealed to come from Chinese LLM teams.

This time, the community detectives were quick to act too. Some tested the tokenizer (the model's word splitter, basically a fingerprint), others watched for API error codes, and eventually they found a Z.AI-specific error format — pointing the biggest suspicion toward Zhipu's GLM-5.3 variant. But as of writing this, no one has come forward to claim it.

Real-world feedback is pretty split too. Some people used it in Claude Code to generate a complete starship simulation page in one shot, commenting "this model overthinks, but the output can definitely hold its own." Meanwhile, developer Bindu Reddy tested it and straight-up said it "spins its wheels badly," with capabilities close to models from two generations ago. The benchmark scores are even more contradictory — one third-party test gave it an average 80% score, while another ranked it dead last. My advice: don't trust any single source on this, just try it yourself and see what's real.

How I Integrated It (Three Steps)

As someone who deals with Hermes Agent every day, of course I had to get it hooked into my workflow ASAP. The whole process was simpler than I expected:

Step one, go to the OpenRouter website, register, and generate an API Key.

Step two, write the Key into Hermes's .env config file.

Step three, change the default model to stealth/ox-alpha in config.yaml, restart, and you're good to go.

# ~/.hermes/.env
OPENROUTER_API_KEY=***

# ~/.hermes/config.yaml
model:
  provider: openrouter
  default: stealth/ox-alpha

A 1M context paired with an Agent is genuinely comfortable. Normally, when I have Hermes analyze a big project, I have to carefully budget tokens just to feed the code in. Now I can just dump the entire repository in and chat away without worrying about blowing up the context window.

Before You Freeload, Remember These Two Things

First, the window is short. The free preview started on August 20th, and officially it only lasts about a week. Pricing and rate limits after that haven't been announced yet. The previous stealth models all got unmasked within a few days, and OxAlpha's true identity is probably about to be revealed too — if you want to try it, move fast.

Second, don't feed it sensitive data. The platform promises zero data retention during the preview period, but that's just a unilateral claim from the vendor, not independently verified. Passwords, personal privacy, company proprietary code — no matter what, never put those into an anonymous model.

I always tell my friends: freeloading is fine, but keep your brain switched on.

It's a bit nostalgic thinking about it. A year ago, we were wincing at the sky-high bills for million-token contexts. Now a model that won't even name itself has just flattened that barrier and handed it over for free.

In 2026, are you still stressing over token costs for long contexts? Let's chat in the comments.