Is Union Alpha Free? Access, Rate Limits, and Real Cost on OpenRouter
Union Alpha is free on OpenRouter right now, but 429 and 502 errors, slow token speeds, and low cache hit rates change the real cost math.
Is Union Alpha actually free right now?
Yes. Union Alpha, a mystery multimodal model that appeared on OpenRouter, is currently available at no cost. It’s also being hosted through Cloudflare. But “free” comes with caveats that matter if you’re planning to build anything serious on top of it: rate limiting, inconsistent throughput, and cache hit rates far below what comparable models achieve. None of that costs you money today, but it costs you time, and if the free window closes or usage-based pricing kicks in later, the token volume this model burns through could get expensive fast.
TL;DR
- Union Alpha is free to use on OpenRouter today, with no confirmed pricing yet if or when that changes.
- Rate limiting is the main practical obstacle, with testers hitting 429 (rate limited) and 502 (server error) responses, plus timeouts, during heavy use.
- Token throughput swings wildly depending on server load, ranging from around 5-6 tokens per second up to 40-50 tokens per second.
- Cache hit rates run low, typically 65-80%, which is well behind DeepSeek models running in comparable harnesses that can hit 99.9% cache hits.
- A single generated project can burn tens of millions of tokens, one first-person shooter game generation used about 50 million tokens with only roughly 68% cached.
- The model reportedly beats GPT 5.6 Sol and Opus 5 on benchmarks like Terminal Bench 2.1 and Sweep Bench Verify, according to OpenRouter’s CEO Alex Atallah, though independent verification is still limited.
- Nobody has confirmed who built it, and there’s public back-and-forth over whether it’s Chinese in origin, with signals pointing more toward a Western, possibly OpenAI-linked, model.
What is Union Alpha, exactly?
Union Alpha is a multimodal model (text and image input, text output) that surfaced on OpenRouter without an official announcement of its creator. It supports a context window of up to 262,000 tokens and is positioned for research, coding, and agentic workflows rather than general chat or creative writing. Its origin is unconfirmed. Speculation online initially pointed to a Chinese lab, but Alex Atallah, OpenRouter’s CEO, publicly pushed back on that framing, and behavioral clues, like how it performs specifically well inside the Codex harness, have led some testers to suspect it’s an OpenAI model or trained on OpenAI-adjacent data.
The only public benchmark data point referenced so far comes from Artificial Analysis, showing Union Alpha outperforming GPT 5.6 Sol and Anthropic’s Opus 5 on Terminal Bench 2.1, and beating GPT 5.6 Sol on Sweep Bench Verify too. If the eventual pricing matches what’s been suggested, that combination of benchmark performance and low cost would put it in a strong position on the cost-versus-performance frontier. But benchmark wins on two tests don’t guarantee consistent real-world performance, and testing so far shows the model’s usefulness varies a lot by task.
Why do rate limits matter more than the price tag?
Because “free” doesn’t mean “usable whenever you want.” Testers report running into three distinct failure modes when hitting Union Alpha through the OpenRouter API: 502 server errors, 429 rate-limit errors, and outright timeouts. These aren’t rare edge cases, they show up regularly enough that generating something as involved as a first-person shooter game took a couple of hours mostly because of rate-limit delays, not because the model itself was slow to reason.
Token throughput is the other variable. Depending on how loaded Union Alpha’s servers are at any given moment, generation speed can range from a sluggish 5-6 tokens per second to a much more usable 40-50 tokens per second. That’s roughly an 8x swing based purely on server conditions, which makes it hard to predict how long any given task will take. If you’re building a workflow that depends on Union Alpha finishing within a specific time window, that unpredictability is a bigger practical constraint than the price being zero.
How does the cache hit rate affect real cost?
This is the detail that matters most if Union Alpha ever moves to paid pricing. Cache hit rate measures how often a model can reuse previously processed context instead of reprocessing it from scratch, and it’s a major lever in the actual cost of running these models at scale. High cache hit rates mean cheaper, faster generations. Low cache hit rates mean you’re paying (or would pay) for a lot of redundant token processing.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Union Alpha’s cache hit rate sits around 65-80% in typical use. Compare that to DeepSeek models running in a DeepSeek-native harness, which can hit cache rates up to 99.9%. That gap is significant. In one documented example, a single generated project (the first-person shooter game) consumed about 50 million tokens, with only around 68% of those cached. Even at a hypothetical low per-token price, a cache hit rate that low means a much larger share of tokens get billed at full rate than you’d see with a more cache-efficient model. If Union Alpha transitions to paid access without improving cache efficiency, the “cheap on paper” pricing could translate into real bills that look a lot less cheap in practice.
Is Union Alpha worth using right now?
It depends heavily on the task and the harness you run it in. Testing showed Union Alpha performing strongly on agentic coding tasks, particularly when run through a verification loop where the model checks its own visual or code output before moving forward. It also showed solid results on things like automated machine learning model training (time series prediction for predictive maintenance, crop density prediction), visual reasoning over images (accurately counting basketball players in a photo, excluding referees), and spatial tasks like arranging simulated crowds into readable text shapes.
Its performance also varies notably by harness. Testing across roughly 10 different agent harnesses found Union Alpha performing comparatively well inside Codex, better than most open models tested there previously, while other harnesses produced more mixed results. That harness dependency is one of the strongest signals that this isn’t just a raw model quality story, it’s also about how well your tooling extracts good behavior from it.
Where it seemed weaker: bounding box precision in object detection tasks (it could correctly identify objects and even read numeric values off technical drawings, but the boxes it drew often didn’t line up correctly), and general creative writing, where testers found it noticeably less compelling than for coding or agentic work.
The practical takeaway: Union Alpha looks like a strong sub-agent or coding-and-research tool right now, not a general-purpose orchestrator or creative writing model. And its “free” status should be treated as a current condition, not a permanent one, especially given how token-hungry some of its agentic workflows can be.
Frequently Asked Questions
Is Union Alpha free on OpenRouter?
Yes, as of now Union Alpha is available for free through OpenRouter, and it’s also hosted via Cloudflare. There’s no confirmed information yet on whether or when that will change to paid pricing.
Why do I keep getting 429 or 502 errors with Union Alpha?
Because it’s a free, high-demand model, its servers can get overloaded. Users have reported 429 (rate limited), 502 (server error), and timeout responses regularly during heavier or longer-running tasks.
Who actually built Union Alpha?
It’s unconfirmed. Early speculation suggested a Chinese origin, but OpenRouter’s CEO publicly disputed that claim, and some behavioral evidence (strong performance specifically in the Codex harness) has led testers to suspect ties to OpenAI, though this remains unverified.
Does Union Alpha support images?
Yes, it’s multimodal on input, accepting both text and images, which allows it to do direct visual analysis and reasoning over photos, diagrams, and other image content. Output is text only.
Is Union Alpha good for creative writing?
Based on testing so far, no. It performs best on agentic coding, research tasks, and structured visual generation, but comes across as noticeably weaker for open-ended creative writing compared to its coding capabilities.
