Union Alpha Arrives as a Free Stealth Coding Model With No Named Maker

On September 16, 2026, a model called Union Alpha appeared on OpenRouter and inside OpenCode with a large context window, image input, tool calling, a zero-data-retention promise and a price of nothing. It arrived without the thing every other frontier release leads with: the name of the laboratory that built it.

Related reading: SpaceXAI Ships Grok 4.7 for Coding at Grok 4.6 Prices, With Vendor Benchmarks to Match · The Week in Tech: The Buildout Meets Its Balance Sheet.

The model is listed as stealth/union-alpha. Its provider field reads simply "Stealth." OpenRouter, which routes requests to it, describes Union Alpha as a multimodal model "built for research, coding, and agentic workflows" and says it delivers "frontier-level performance across a broad range of general-purpose tasks." Crypto Briefing reported the launch as a collaboration between OpenCode and OpenRouter, live on September 16 and free "during an initial preview period lasting roughly one week."

An anonymous model is not new for this platform. OpenRouter has a running habit of hosting "stealth" releases, and OpenCode distributed a similar one, Omen Alpha, earlier. What is different this time is the combination: a free preview, a 262,144-token context window, image support, and a set of benchmark claims that, so far, come mostly from the community rather than a vendor.

What Union Alpha actually is

Strip away the mystery and the specification sheet is concrete. Union Alpha accepts text and images and returns text. It carries a 262,144-token context window with a maximum output of 131,072 tokens, according to its OpenRouter listing. The API accepts tool calling, tool choice, response format and streaming, which is the minimum an agent needs to inspect a repository, run a command, read the result and decide what to do next.

OpenCode's own documentation lists Union Alpha as not used for training and with zero days of data retention on its Go route. That is a stronger privacy posture than many hosted coding models offer, and for developers working on private repositories it is the most consequential line on the page.

The absences matter just as much. There is no parameter count, no architecture note, no published weights and no way to run the model yourself. Union Alpha is not a family a developer can inspect; it is an endpoint a developer can call.

The context window, in practice

At a quarter of a million tokens, the context window is the headline hardware number. It is roughly enough to hold 200,000 words, or a large repository and its documentation in a single pass, as Crypto Briefing noted. For agentic coding that capacity changes what a single session can attempt. A model that can see an entire service at once does not need a retrieval layer to decide which file matters.

Capacity is not the same as quality, though, and the honest caveat is that a large window says nothing about whether a model can find the right detail from thirty thousand tokens back. Long-context retrieval is a separate skill from long-context storage, and the community testing so far has put more weight on task completion than on needle-in-a-haystack recall.

Why image input changes the coding workflow

The second capability worth dwelling on is image input. Text-and-code models have been able to read diffs for years, but a large share of real debugging is visual. A broken layout, a misaligned component, a design mock that does not match the build, a chart with the wrong axis: these arrive as screenshots, not as prose. A model that accepts an image can be handed the rendered page and the failing source at the same time and asked to reconcile them.

For frontend and product work this is a genuine shortcut. Instead of describing a visual bug in words and hoping the description survives translation into code, a developer can attach the screenshot and let the agent connect what it sees to the implementation. It is the kind of capability that looks minor in a specification table and turns out to be the reason an agent is useful on a Tuesday afternoon. The current metadata describes text output only, so Union Alpha reads images rather than generating them; it is a multimodal reasoner, not a media tool.

The benchmark picture is community-shaped

Because no vendor published a scorecard, the performance story is being assembled by third parties and developers. The measurements that exist point in different directions, and each one needs its caveats attached.

Benchable, an independent model tracker, reports that Union Alpha achieved a 100 percent success rate across eight benchmarks it ran, with coding at 94.0 percent accuracy (83rd percentile), reasoning at 98.0 percent (90th percentile) and mathematics at 94.9 percent (79th percentile). Its weaker areas were instruction following, at 76.0 percent, and raw speed, where it landed in the 6th percentile. In plain terms: accurate on the tasks it finishes, slow to finish them.

A separate evaluation circulated at launch by BuildFastWithAI recorded 23.14 out of 40 across four coding projects, at an average cost of about three cents per prompt, and cited a developer report of roughly 51 percent on Terminal-Bench v4 at about $1.50 per task. That Terminal-Bench figure is the kind that should not be quoted without air quotes; its full methodology was not published, so it is a claim, not a comparable score.

The disagreements are not contradictions so much as different tests. Benchable measures accuracy on bounded questions. The coding snapshots measure whether an agent finishes a real task. Those two things diverge often, and for anyone evaluating a model for agent work, the second is the one that pays the bills.

The stealth-model playbook

Union Alpha did not invent its own release strategy. OpenRouter has become a regular venue for models that appear under an "Alpha" name with no developer attached, and the pattern is consistent: a capable endpoint, a free or near-free preview, a deliberately thin information page, and a surge of community speculation about who is really behind it.

OpenCode shipped Omen Alpha under similar terms in August. Benchable's page lists Omen Alpha as a prior stealth release dated August 20, 2026, with a one-million-token context window and video input. Omen Alpha was later withdrawn from OpenCode's model list. The two models make a useful pair: Omen had the larger window, Union Alpha has the newer benchmark data and the image-and-tools combination that today's coding agents lean on.

The reasons a lab might release this way are not hard to guess. A stealth launch tests real-world demand and behavior before a company commits its name to a model, and it lets a lab gather usage data and head-to-head comparisons without a marketing campaign or a promise about pricing. It also sets up a second announcement: when the model is eventually revealed, it arrives with a week of usage behind it and a name that developers have already been typing. For developers, the trade is straightforward: access to a capable model for free, in exchange for evaluating a black box.

There is a cost to that trade that is easy to overlook. When a model is anonymous, so is its failure mode. If a named lab ships a model that mishandles long context or drops tool calls, the community learns the pattern quickly and the vendor answers for it. A stealth model has no reputation to protect and no channel through which to explain itself. Whatever an individual developer discovers about Union Alpha stays with that developer unless they publish it, which is part of why the community measurements matter more here than they would for a signed release.

Who made it, and why that matters

The provider has not been identified. OpenCode labels the model a stealth model, and registries list the provider as "Stealth" rather than any named company. That has not stopped the guessing. Community theories have pointed at Chinese labs, with Moonshot AI and Zhipu, the maker of the GLM family, surfacing in discussions, as Crypto Briefing reported. None of those connections have been confirmed, and the reporting is explicit that they remain speculation.

For a developer deciding whether to route production work through the endpoint, the gap is not cosmetic. Knowing the lab tells you about model lineage, where the data goes, what the roadmap looks like and what license will apply. Without it, Union Alpha can only be judged as a service: what it does on a controlled test, what it costs after the preview, and what happens to the code you send it.

Two practical points follow. First, the zero-retention promise applies to the OpenCode Go route. Third-party hosts can set their own terms, so the policy of the endpoint you actually call is the one that governs. Second, an anonymous model is a reason to keep secrets, credentials and tool permissions isolated regardless of the retention policy, because those controls protect you even if a policy changes.

What to watch after the free week

The preview is the easy part. The questions that will decide whether Union Alpha matters start when the free window closes.

• Price. No post-preview rate has been published. A one-week free trial is an evaluation opportunity; it is not a production cost baseline.
• Attribution. If a lab claims the model, the questions about lineage, data handling and licensing get concrete answers. If nobody does, the black box stays a black box.
• Consistency. Community reports have flagged context recall and in-OpenCode failures. Whether those are launch-week rough edges or a lasting trait will only show up under repeated use.
• Comparison. The honest test is not Union Alpha against a marketing page. It is Union Alpha against the model you already use, on the same repository, the same prompts and the same timeout policy.

The bottom line

Union Alpha is a genuinely unusual release: a real, capable-seeming endpoint attached to no publicly named maker, handed out free for a week with a privacy posture that would be notable even on a signed product. For a coding agent, the profile is attractive. A quarter-million-token window, image input and tool support cover most of what agentic work asks for, at a price of zero during the preview.

The caution is equally clear. The benchmarks are community measurements with stated limits, the provider is undisclosed, and the long-term cost is unknown. The sensible approach is the one the model itself invites: use the free week to run your own tasks against it, judge it on completed work rather than on the mystery around its name, and keep a documented frontier model in reserve for changes where being wrong is expensive.

Cover image: the official OpenCode brand mark. OpenCode is the platform through which Union Alpha is distributed; the mark is not the model maker's logo, which is not public. Asset from the official OpenCode brand assets (opencode.ai/brand).