OpenRouter quietly listed a free model called Crouton Alpha this week with a full 1 million token context window. Nobody will say who built it, and that's becoming the whole business model.
OpenRouter added Crouton Alpha to its catalog at 01:38 UTC on November 3, under the site's anonymous "Stealth" provider tag. The listing describes it plainly: a reasoning model built to handle the repeated sub-tasks AI agents chew through all day. It has a 1,048,576 token context window and supports up to 944,000 tokens of output. It takes text, image, audio and video as input. The price on every field is zero. OpenRouter is explicit that it did not build the model and does not know, or will not say, who did.
That context window puts Crouton Alpha in the same bracket as Google's Gemini 1.5 and 2.0 Pro, both built around a million-token window, and ahead of what OpenAI and Anthropic ship by default. Say you're building something that has to read a full codebase, a legal filing, or hours of video transcript in one shot. A window that size is the difference between chunking the input and just pasting the whole thing in.
Running a model with a million-token window, multimodal input, and near-trillion-token daily capacity for free costs real money in inference compute. That's exactly the point of a stealth drop. OpenRouter has run this playbook for almost two years now. Quasar Alpha and Optimus Alpha showed up free on the platform in April 2025 and were later confirmed as early test builds of OpenAI's GPT-4.1. Xiaomi's MiMo team has been tied to Hunter Alpha, the trillion-parameter, 1M-context model that surfaced in March 2026. Space Bunny Alpha, listed September 23, 2026, turned out to be MiniMax's M3.1 Flash. Ox Alpha landed in August 2026 with the same 1,048,576-token window Crouton Alpha now carries, and it was never officially claimed. But developers ran tokenizer analysis on its outputs and found a byte-fallback pattern and vocabulary layout nearly identical to Zhipu AI's GLM series. The error payloads matched Zhipu's BigModel cloud infrastructure too.
Crouton Alpha hasn't been fingerprinted yet. Posts tracking the release on X have floated Mistral, an unnamed low-cost Chinese lab, and xAI's Grok line as guesses. None of it is confirmed: anyone telling you otherwise is speculating same as everyone else.
China's GLM-5.3-Flash Just Beat Claude Opus 4.8 on Agent Benchmarks Zhipu AI revealed that its mysterious "Ox Alpha" test model was actually GLM-5.3-Flash, an MIT-licensed open-weight model that beats Claude Opus 4.8 on key agent benchmarks while costing roughly 42 times less. The release piles more pricing pressure on US labs just as agentic coding tools become the biggest driver of token spend. - GLM-5.3-Flash beats Claude Opus on agent benchmarks - China's GLM-5.3-Flash model pricing and performance comparison
The pattern itself is the story.
A lab gets a free period of real-world load testing, a flood of benchmark chatter, and a name nobody can trace back to a press release. All the while, competitors scramble to figure out who just undercut them on price and context length at once. Free and anonymous is cheaper than a launch event, and it generates more organic buzz than one too.
For the developers actually using it, the practical question is whether the million tokens are real or just a number on a spec page. Early chatter puts Crouton Alpha's throughput at a median of roughly 124 tokens per second, with no official benchmark scores published by the provider or by OpenRouter itself. That's the catch with every stealth release: there's no paper and no model card with eval numbers. There's also no one to hold accountable if the long-context performance degrades the way it has on some rival models once you push past a few hundred thousand tokens. OpenRouter also caps free-tier usage with per-minute and per-day rate limits, so "free" doesn't mean unlimited. Builders testing it for production workloads will hit that ceiling fast. None of that has slowed adoption. Owl Alpha, an earlier stealth entry from April 2026, climbed to the seventh most-used model on the entire OpenRouter platform within weeks of its listing, without its backers ever confirming who made it. If Crouton Alpha follows that curve, whoever built it will get what they wanted: a few weeks of unfiltered, high-volume feedback from real agentic workloads, at the cost of giving the compute away.
Also read: AI Infrastructure Stocks Sink as Democratic Midterm Sweep Threatens Data Centers • AI Is Quietly Ending the Fiverr and Upwork Era of Gig Work • Supermicro posts a $19.9 billion quarter and the AI bubble doubters go quiet
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.