Union Alpha Pricing
Free today, roughly $0.65 per coding task at the anticipated rate, and still far below the $6.50 to $11.80 that comparable models charge.
Pricing breakdown
| Item | Input | Output | Note |
|---|---|---|---|
| Preview (current) | $0 | $0 | Free while the preview runs; no published end date |
| Anticipated rate | $0.50 / 1M | $1.50 / 1M | Reported anticipated pricing, not officially confirmed |
| Estimated per task | — | ~$0.65 | Typical coding task at the anticipated rate |
| Frontier comparison | — | $6.50 – $11.80 | What GPT and Opus-class models charge for the same task |
How it stacks up
Even at the anticipated post-preview rate, Union Alpha would still undercut frontier competitors by a wide margin for comparable coding output.
| Model | Per task | Note |
|---|---|---|
| Union Alpha | ~$0.65 | Free during preview; anticipated $0.50/$1.50 per 1M |
| GPT-class frontier | $6.50 – $11.80 | Comparable SWE-bench resolution rate |
| Opus-class frontier | $6.50 – $11.80 | Comparable SWE-bench resolution rate |
| GLM-5.3-Flash | ~$0.30 | Cheaper, but ~33% on Terminal-Bench v4 vs Union Alpha's ~52% |
What This Costs in Practice
Per-token rates are hard to reason about. Here is what the anticipated pricing works out to across three realistic usage patterns, assuming a typical coding task consumes roughly 40,000 input tokens and 8,000 output tokens.
| Usage pattern | Tasks / month | Estimated cost | Same work on a frontier model |
|---|---|---|---|
| Side project | ~50 | ~$33 | $325 – $590 |
| Daily driver | ~300 | ~$195 | $1,950 – $3,540 |
| Agentic pipeline | ~2,000 | ~$1,300 | $13,000 – $23,600 |
These are estimates built on the anticipated rate, not a published price sheet. Real token consumption varies enormously with prompt design, how much repository context you attach, and how many turns an agent takes before it converges.
Planning for the End of Free
The preview has no published end date, which is exactly why you should assume it ends. Stealth listings on OpenRouter have historically run free for a limited window and then either switched to paid rates or disappeared entirely when the model relaunched under a real name.
Three things worth doing while access costs nothing. First, benchmark it against whatever you currently pay for on your own workload — vendor benchmarks rarely match real tasks, and this is a free opportunity to get honest numbers. Second, keep your integration behind an abstraction so swapping the model string is a one-line change rather than a refactor. Third, if the quality holds up on your evaluation, budget for the anticipated rate now rather than being surprised later.
The upside case is straightforward: even at full anticipated pricing, the cost advantage over comparable-scoring models is roughly tenfold. The downside case is that the listing vanishes overnight with no vendor to call. Both are live possibilities, so treat this as a cheap experiment rather than infrastructure.