Union Alpha vs Fable 5.1 vs GPT-5.6 Terra: Stealth Model
The short answer
Union Alpha is the September 2026 stealth model: anonymous, free for about a week, and built for agentic coding. Early numbers put it a point behind the best frontier coders on a software-engineering benchmark at roughly a tenth to a twentieth of the cost per task. Claude Fable 5.1 is the verified top of the coding stack and costs accordingly; GPT-5.6 Terra is the mid-tier OpenAI model whose price Union Alpha is expected to match. Use Union Alpha to experiment, not to ship.
| Union Alpha (stealth) | Claude Fable 5.1 | GPT-5.6 Terra | |
|---|---|---|---|
| Operator | Undisclosed | Anthropic | OpenAI |
| Available since | September 16, 2026 (OpenRouter, OpenCode) | September 3, 2026 (GA) | Summer 2026 |
| Context / max output | 262K / 131K tokens | 1M / 128K | 1M-class |
| Input types | Text, images; tool calling | Text, images; tools | Text, images; tools |
| Price (per MTok) | Free during preview; expected “Terra-class” | $10 / $50; cache reads $0.25 | $2 / $12 |
| Coding signal | ~73% on a SWE-style chart at ~$0.65/task (unverified) | AA Intelligence Index 65.6 (#1); $9.18/task at max effort on AA’s coding harness (70.4 score) | Mid-tier; not on the top of coding leaderboards |
| Data policy | Zero retention on prompts (vendor claim, anonymous) | Contractual; no training on API data by default | Contractual; no training on API data by default |
| Day-zero usage | ~1.96B tokens on OpenRouter | n/a | n/a |
What we know about Union Alpha
The model appeared on September 16, 2026 on OpenRouter as stealth/union-alpha and inside OpenCode, the open-source terminal coding agent, which offered it free for a week with no data training and image support. OpenCode’s post described it as “built for agentic coding.” Specs published with the listing: 262K context, 131K max output, multimodal input, tool calling. OpenCode advertised 5 trillion tokens per day of capacity, which is a serious infrastructure commitment for a product with no named owner, and OpenRouter traffic hit about 1.96 billion tokens on launch day.
Community benchmarking placed it “close to Fable 5 level at GPT-5.6 Terra pricing.” A widely shared chart from a software-engineering benchmark had Union Alpha at ~73% versus ~74% for the two leading GPT and Opus-class models, at ~$0.65 per task against $6.50 and $11.80. In a hands-on Docker debugging test (Postgres, Flask, Redis, mixed Python/JavaScript), it correctly diagnosed and fixed a subtle scoring bug that averaged all-time history into a seven-day metric. Testers noted it prefers reasoning and testing in Python even in JavaScript-heavy repos, and that free-tier throttling cut evaluations short. One report claimed it used about 3x fewer tokens than Ox Alpha on a 3D-scene task.
When probed through OpenCode’s subagent tool, the model refused to reveal its identity, by design.
How Claude Fable 5.1 compares
Fable 5.1 went GA on September 3, 2026 at the same $10 / $50 per million tokens as Fable 5, but with cache reads cut 75% to $0.25, which Anthropic estimates makes typical workloads ~25% cheaper and agentic workloads up to ~45% cheaper. It leads the Artificial Analysis Intelligence Index at 65.6 and tops AA’s coding-agent harness at max effort (70.4) — at $9.18 per task, the most expensive entry on that chart. It is available day-one on AWS, Google Cloud, Azure and Anthropic’s API, with a documented data policy and a company to hold accountable. If Union Alpha’s ~73% at ~$0.65 holds up, it is a genuine value story; if it does not, Fable 5.1 remains the safe pick for hard agentic coding.
How GPT-5.6 Terra compares
Terra is OpenAI’s mid-tier model in the GPT-5.6 line, priced at $2 input / $12 output per million tokens since the July 30, 2026 20% cut (down from $2.50/$15). It sits well below Sol ($4/$20) and Astra ($10/$50) on capability; it is the model to compare Union Alpha against on price, not on frontier coding quality. Reference cost for a 30K-in / 5K-out task: about $0.12 on Terra versus $0.55 on Fable 5.1.
The risk that does not show up in benchmarks
The stealth-model pattern is now familiar: drop something capable, make it free, gather usage before anyone forms a critical opinion. Ox Alpha (August 20, 2026) was revealed in September to be Z.ai’s GLM-5.3 Flash, served on roughly 100,000 domestic Chinese accelerators. Union Alpha’s zero-retention claim is better than Ox Alpha’s silence, but it is still an unverifiable promise from an unnamed party — no jurisdiction, no subprocessor list, no contract.
Rules of thumb:
- Public repos, synthetic tasks, benchmarks: fine. This is what the free week is for.
- Proprietary code, secrets, customer data: no. Wait for the named launch and a data-processing agreement.
- Do not build a pipeline on it. The endpoint will change or vanish when the vendor announces; OpenRouter had already surfaced a successor stealth listing (
unbiased/pareto) by September 17. - Measure tokens per task, not price per token. Union Alpha’s reported efficiency, if real, is the durable advantage.