Union Alpha Benchmarks
Results from Artificial Analysis's Terminal-Bench v4.0 — a 149-model evaluation of terminal-based coding and agentic capabilities.
At a glance
Bottom line: Union Alpha occupies the most attractive quadrant of the chart — it delivers mid-to-high terminal capability at a cost far below frontier models. On the Pareto frontier between flash-tier and Opus-tier.
Where Union Alpha sits on the chart
Reconstructed from Artificial Analysis's Terminal-Bench v4.0 scatter plot (28 of 149 models shown). X-axis = score (higher is better), Y-axis = cost per task in USD (log scale, lower is better). The Pareto line marks the efficient frontier — any model above or to the left of it is dominated by a cheaper and/or better alternative.
Terminal-Bench v4.0 — Score vs Cost per Task
Pareto frontier analysis
Models Union Alpha beats on cost
Every model on the high-end frontier — Claude Opus 5, GPT-6 Astra, Claude Fable 5.1 — costs significantly more per task while delivering a score within the same order of magnitude. For terminal-based agentic workflows where you run hundreds of tasks, the per-task cost gap compounds quickly.
Models Union Alpha beats on capability
Budget-tier models like GLM-5.3-Flash and Gemini 3.5 Flash-Lite are cheaper per task, but their terminal proficiency scores are materially lower — they are optimised for speed and cost, not terminal reliability. Union Alpha bridges the gap: it costs marginally more than flash models but delivers a meaningful capability jump.
The price caveat you need to know
The $4-5/task figure on the chart is labelled "anticipated pricing" by TokenRa. Union Alpha is currently free (see our API page for connection details). When it transitions to paid tier, the benchmark suggests pricing will land in this range — making it a mid-cost, mid-to-high-capability option.
How Union Alpha stacks up
| Model | Score | Cost / Task | Tier | Note |
|---|---|---|---|---|
| Union Alpha | ~50–60% | ~$4–5 (anticipated) | Mid | Currently free on TokenRa |
| GLM-5.3-Flash | Lower | Lower | Flash | Optimised for speed, not terminal reliability |
| Gemini 3.5 Flash-Lite | Lower | Lower | Flash | Best for high-throughput, low-complexity tasks |
| GPT-6 Astra | Higher | Higher | Frontier | Full GPT-6 family capability, higher cost |
| Claude Fable 5.1 | Higher | Higher | Frontier | Anthropic mid-to-high tier |
| Claude Opus 5 | Highest | Highest | Frontier | Top-tier, premium pricing |
Score and cost ranges are estimated from the Artificial Analysis chart (log-scale axes). Exact figures are not published; positions are approximate.
What this benchmark measures
Terminal-Bench v4.0
Terminal-Bench is an internal evaluation framework by Artificial Analysis that tests models on terminal-based coding and agentic tasks. Unlike general-purpose benchmarks (MMLU, HumanEval), it specifically measures how well a model performs in a command-line environment — running commands, editing files, navigating file systems, and completing multi-step agentic workflows.
What the score means
The score (vertical axis) reflects task success rate: a model scoring 50% successfully completes approximately half of the terminal tasks it's given. The cost (horizontal axis, log scale) is the estimated USD cost per task at the model's anticipated pricing.
Data source & limitations
- Source: Artificial Analysis (retrieved September 2026)
- Chart shows 28 of 149 evaluated models — the full dataset is not publicly available
- "Anticipated pricing" means Union Alpha is currently free; this is TokenRa's estimate for future paid tier
- Benchmark results may change as the model is updated by the Stealth provider
- Union Alpha is a stealth model — the provider is anonymous and model capabilities may shift without notice
Frequently asked questions
What is the Union Alpha Terminal-Bench score?
Union Alpha scores approximately 50-60% on Terminal-Bench v4.0 from Artificial Analysis, at an anticipated cost of ~$4-5 per task. This places it on the Pareto frontier between cheaper flash models and high-end frontier models.
How does Union Alpha compare to other models on benchmarks?
Union Alpha outperforms budget-tier models such as GLM-5.3-Flash and Gemini 3.5 Flash-Lite in terminal capability while costing less than frontier models such as GPT-6 Astra or Claude Opus 5. It occupies a mid-range sweet spot on the cost-performance Pareto curve.
Is Union Alpha currently free or paid?
Union Alpha is currently free to use on TokenRa. The benchmark pricing of $4-5 per task is listed by TokenRa as "anticipated pricing" for when it moves to paid tier.
What benchmark is used for the Union Alpha results?
The results come from Artificial Analysis's Terminal-Bench v4.0 (Internal-Terminal-Benchmark), which evaluates models on terminal-based coding and agentic tasks across 149 models.
Why should I trust these benchmark numbers?
Artificial Analysis is a widely-cited independent evaluator. However, terminal benchmarks measure one narrow capability (CLI agentic performance) and do not reflect chat quality, reasoning depth, or instruction-following. Use this alongside other evaluations for a full picture.
Will Union Alpha's benchmark scores change?
Yes. Stealth models are updated silently by their anonymous providers. A model called "Union Alpha" today may have different capabilities next week. Check TokenRa's model page and Artificial Analysis for the latest results.