Union Alpha is the new stealth coding model that appeared on OpenRouter and OpenCode on 16 September 2026, free to use, with nobody willing to say who built it. It landed with a 262,144-token context window, image input, tool calling, and a preview period that both platforms describe as roughly one week.
The pattern is familiar by now. A frontier-grade model shows up under a codename, gets routed through an anonymous provider slot, collects a few million tokens of real developer usage, and either gets a name later or quietly disappears. What makes this launch worth reading closely is not the marketing copy on either site. It is the machine-readable metadata underneath it, which the two platforms publish in full and which does not entirely agree with itself.
This article works through what is verifiable: the exact specification as OpenRouter’s own API reports it, the cost of the free week measured against what the same work costs on paid models, a direct contradiction between OpenCode and OpenRouter on data retention, and what the routing configuration hints about the model’s origin without proving it. Everything below was checked against the live APIs on 17 September 2026 rather than taken from the launch posts.
Table of contents
- What Union Alpha Is and Where It Came From
- Union Alpha Specifications Verified Against the API
- The Free Week: What Zero Cost Actually Buys
- The Data Retention Contradiction Between the Two Platforms
- What the Routing Metadata Says About Who Built Union Alpha
- Union Alpha Versus Ox Alpha, the Last Stealth Model
- How to Run Union Alpha Today
- What to Watch Before You Put Union Alpha in a Workflow
- Why Labs Keep Launching Models in Stealth
- Frequently Asked Questions About Union Alpha
- References and Further Reading
What Union Alpha Is and Where It Came From
A stealth model is a production model served under a codename by a provider who has not been publicly identified. OpenRouter runs a dedicated slot for them, and this is the second high-profile entry in under a month.
The launch timeline, to the minute
OpenRouter’s model catalogue records the creation timestamp for stealth/union-alpha as 1789569723, which is 14:42:03 UTC on 16 September 2026. OpenRouter’s announcement post followed six minutes later at 14:48 UTC, and OpenCode’s own post went out at 14:52 UTC. The model then appeared in OpenCode’s Zen gateway catalogue at 19:56 UTC the same day, about five hours behind the OpenRouter listing.
What a stealth model actually means
OpenRouter’s disclaimer is explicit about the arrangement: “This stealth model is developed and operated by a third-party model provider.” OpenRouter is the router, not the owner, and it says so on the page. The provider name in the API response is simply “Stealth”, the quantization field reads “unknown”, and there is exactly one endpoint, so there are no routing decisions and no provider comparison to make.
The two places you can reach it
Only two surfaces carry it. On OpenRouter it is the model ID stealth/union-alpha. On OpenCode it is union-alpha, listed in the Zen pricing table as “Union Alpha Free”. There is no Hugging Face repository, no weights, no model card, no paper and no benchmark disclosure from whoever built it — which is the entire point of a stealth launch.
The promotional framing on each platform
OpenRouter’s post described it as “A multimodal model for research, coding, and agentic workflows” with “Frontier-level general-purpose performance”. OpenCode’s was blunter and more developer-facing: “free for the next week”, “no data training”, “built for agentic coding”, “supports images”. Neither post named a lab, a parameter count, a training cutoff, or a single benchmark score.
Union Alpha Specifications Verified Against the API
The launch posts round the numbers. OpenRouter’s public API does not, so the table below reads from /api/v1/models and the model’s endpoint detail response rather than from any announcement.
| Field | Reported value | What it means in practice |
| Model ID | stealth/union-alpha | Provider slot is literally named “Stealth” |
| Context length | 262,144 tokens | Exactly 256K; the post rounds it down |
| Max completion | 131,072 tokens | Half the context can be output |
| Modality | text+image to text | Screenshots in, no image or audio out |
| Price | $0 input, $0 output | Zero on both platforms during preview |
| Tokenizer | “Other” | Not matched to a known public family |
| Moderated | false | No OpenRouter-side moderation layer |
| Uptime, last 24h | 99.986% | Serving infrastructure is production-grade |
| Knowledge cutoff | null | Not disclosed by the provider |
The context window is the headline number
262,144 tokens is roughly 200,000 words of input, which covers a mid-sized repository in a single call without retrieval or chunking. That is the specification agentic coding actually needs, because an agent that has to re-read files between turns burns most of its budget on navigation. Union Alpha can hold the map and the territory at once.
What the parameter list quietly tells you
The supported parameter list is short: max_tokens, temperature, top_p, tools, tool_choice and response_format. There is no reasoning or include_reasoning parameter, no seed, no stop, no top_k, no logprobs and no frequency or presence penalties. The absence of a reasoning parameter is the notable one — it means Union Alpha either does not expose a separate thinking channel or the provider has chosen not to surface it through the router.
Tool calling is narrower than the tweet suggests
OpenRouter advertised “Tool calling”, and the endpoint detail confirms tools are supported. But the supports_tool_choice object shows only auto set to true. The none, required and named-function modes all report false. If your agent framework forces a specific tool or disables tools per turn, that control is not available here, and the DevOps pipelines built around strict tool routing will need a fallback path.
The listing has no expiry date set
Both platforms describe a week-long preview, but the OpenRouter catalogue entry carries an expiration_date of 2098-12-31 — the placeholder value used for models with no scheduled removal. So the free week is a commercial commitment made in a social post, not a technical one written into the listing.
The Free Week: What Zero Cost Actually Buys
Free is easy to say and hard to size. The cleanest way to measure what is on offer is to price the same work against the paid models on the same gateway, since OpenCode publishes its full rate card.
One full context round trip, priced across the Zen rate card
Take a single maximal call: fill the 262,144-token context and generate the 131,072-token maximum output. That is 0.262144M input tokens and 0.131072M output tokens. Multiply by each model’s published per-million rate on OpenCode Zen and the comparison is direct.
The arithmetic behind the top bar: 0.262144 multiplied by $10.00 gives $2.62 of input, and 0.131072 multiplied by $50.00 gives $6.55 of output, for $9.17 in total. Run that call fifty times across a week of agent work and the Claude Fable 5.1 equivalent is roughly $459. That is what the preview is handing out, per developer, and it explains why these launches draw the traffic they do.
Why the preview has no published end date
Neither platform committed to a date. OpenCode’s wording is “free for the next week” and its documentation says “available on OpenCode for a limited time”. OpenRouter’s pricing page simply reads $0 with no countdown. Pricing after the preview has not been disclosed anywhere, which means anyone wiring Union Alpha into a production path is budgeting against an unknown number.
The capacity question nobody answered this time
When the previous stealth model launched, OpenCode published a capacity figure — 100 trillion tokens per day — and described the rate limits as generous enough to call the usage “near unlimited”. No equivalent claim accompanied Union Alpha. The omission may mean nothing, or it may mean the provider is running a tighter serving budget. Either way, plan for throttling rather than assuming the previous generosity carries over.
The Data Retention Contradiction Between the Two Platforms
This is the part of the launch that deserves more scrutiny than it received, because the two places you can use the model make materially different promises about your code.
What OpenCode says
OpenCode’s Zen documentation states it plainly: “Union Alpha Free is a stealth model available on OpenCode for a limited time. Its provider follows a zero-retention policy and does not use your data for model training.” Its launch post shortened that to “no data training”. Most coverage repeated the phrase “zero data retention” on that basis.
What OpenRouter says
OpenRouter’s disclaimer on the same model reads differently: “Prompts and completions for this model may [be] retained by the provider but are not used for training; all other use is governed by the Stealth Model Terms.” That is retention permitted, training excluded — a weaker guarantee than zero retention, for the identical model behind the identical anonymous provider.
| Source | Stated policy | Prompts retained? | Used for training? |
| OpenCode Zen docs | “zero-retention policy” | No | No |
| OpenCode launch post | “no data training” | Not addressed | No |
| OpenRouter model page | “may be retained… not used for training” | Possibly | No |
| The provider | No public statement | Unknown | Unknown |
Which wording should govern your decision
Treat the weaker claim as the operative one unless you are certain which route your request takes. Both statements are second-hand: neither OpenCode nor OpenRouter built the model, so both are relaying a policy an unnamed third party told them. For anything covered by a client NDA, an employment contract or a data protection obligation, that chain of assertion is thin. The safe reading is that Union Alpha may retain what you send it.
What this means for real codebases
If you want to try Union Alpha on work code, the sensible approach is the one you would apply to any unidentified vendor: use it on an open-source repository, a personal project, or a sanitised subset. Save the proprietary monorepo for a model whose operator you can name and whose terms you have actually signed. That is not paranoia about this particular model — it is the baseline for a supplier you cannot identify.
What the Routing Metadata Says About Who Built Union Alpha
Nobody has confirmed the lab. But the launch left one artefact behind that is more interesting than the usual tokenizer guessing, and it sits in OpenCode’s own published endpoint table.
The Anthropic Messages endpoint clue
Every model on OpenCode Zen is listed with the endpoint it serves on and the AI SDK package that talks to it. Union Alpha is routed to https://opencode.ai/zen/v1/messages using @ai-sdk/anthropic — the Anthropic Messages API shape, not the OpenAI-compatible /chat/completions shape that most non-Anthropic models on the gateway use. A provider does not pick that surface by accident; it reflects the API the upstream model actually speaks.
Why that is not proof of anything
Here is where most of the speculation goes wrong. The Messages-API group on Zen is not Anthropic-only. Of the 70 models in that table, 16 route through @ai-sdk/anthropic: eleven Claude models, four Qwen models, and Union Alpha. Alibaba’s Qwen tier serves on the Anthropic surface too, because shipping an Anthropic-compatible endpoint has become a standard way for any lab to slot into coding agents built for Claude. So the clue narrows the field to labs that ship that surface. It does not name one.
The other fingerprints point in different directions
OpenRouter reports the tokenizer as “Other”, meaning it does not match a recognised public family — which is what you would expect from either a genuinely new tokenizer or a deliberately obscured one. The 262,144-token context is a common size across several Chinese open-weight families. The lack of an exposed reasoning channel is unusual for a 2026 frontier release, since most coding models of this generation are post-trained with reinforcement learning and advertise the thinking budget as a feature. None of these signals agree with each other, which is the honest summary.
What the speculation got wrong last time
There is a recent precedent worth remembering. When the previous stealth model appeared, developer analysis pointed at Xiaomi’s MiMo team on tokenizer grounds, and separately at Z.ai on specification grounds. The answer turned out to be Z.ai, confirmed only after days of confident guessing in both directions. The lesson applies directly to Union Alpha: the metadata narrows the field, and the field is still wide.
Union Alpha Versus Ox Alpha, the Last Stealth Model
The obvious comparison is the stealth launch four weeks earlier, and the two are not the same proposition despite the matching branding.
| Attribute | Union Alpha | Ox Alpha |
| Launched | 16 September 2026 | 20 August 2026 |
| Context window | 262,144 tokens | 1,048,576 tokens |
| Max output | 131,072 tokens | 131,072 tokens |
| Input modalities | Text, image | Text, image, video |
| Stated capacity | None published | 100T tokens per day |
| Retention claim | Contradictory across platforms | “Zero Data Retention” |
| Lab identified? | No | Yes, later — Z.ai |
| In OpenRouter catalogue today | Yes | No longer listed |
Smaller context, narrower modality
Union Alpha’s context window is exactly a quarter of its predecessor’s: 262,144 against 1,048,576. It also drops video input, accepting only text and images. On raw specification, this is the smaller model of the two, which makes the “frontier-level” claim a statement about quality rather than scale.
The attention gap is the real story
OpenCode’s launch post for the earlier model drew 8,524,810 views and 15,834 likes. Its Union Alpha post, read on 17 September, sat at 963,207 views and 7,724 likes — about 11% of the reach, at roughly half the likes. OpenRouter’s own post added 504,091 views and 1,917 likes. The novelty of the stealth format is wearing off faster than the launches are arriving.
Ox Alpha has already left the catalogue
A check of OpenRouter’s live model list on 17 September 2026 returns 444 models, and stealth/ox-alpha is not among them. The previous stealth model was pulled from the catalogue once its preview ended, even though it is still described on the stealth landing page. That is the most concrete answer available to “what happens when the free week ends”, and anyone building on Union Alpha should assume the same disappearance is the default outcome.
How to Run Union Alpha Today
Both routes take minutes. Neither requires any special access, waiting list or approval.
Through OpenRouter
Point any OpenAI-compatible client at OpenRouter’s chat completions endpoint with your OpenRouter API key and set the model to stealth/union-alpha. Because the listed price is $0 on both input and output, the calls do not draw down account credit during the preview. Only one provider serves the model, so there is no routing preference to configure and no fallback ordering to think about.
Through OpenCode
In OpenCode’s terminal agent, sign in to Zen, run /connect, then select the model with the identifier opencode/union-alpha. The Zen pricing table lists it as free on input, output and cached reads. This is the route that most of the launch-week testing is going through, because it puts the model straight into a working coding agent rather than a chat box.
What to actually test it on
Give it work that uses the specification. Load an entire repository and ask for a cross-file refactor. Paste a screenshot of a broken UI and ask for the CSS fix, since image input is supported. Run a multi-step tool-calling task with tool_choice left on auto, which is the only mode available. Comparing it against your current assistant on the same tasks is the only benchmark that matters here, given that no official scores exist — our own comparison of AI coding assistants is a reasonable baseline to test against.
Keep a record of what you find
Stealth previews are short and the community evaluation is the only evaluation. If Union Alpha does something unusual — an unexpected refusal, a distinctive formatting habit, a specific benchmark result — write it down while the model is still reachable. That record is what will let you recognise it when it resurfaces under a real name, which is exactly how the previous model was eventually pinned to its lab.
What to Watch Before You Put Union Alpha in a Workflow
Free and fast is a good reason to experiment. It is not on its own a reason to depend on something, and there are four specific flags on this listing.
The moderation flag is set to false
OpenRouter’s endpoint data reports is_moderated: false. There is no router-side safety filter in front of Union Alpha, so whatever alignment behaviour exists is whatever the anonymous provider built into the model. For a coding tool that is usually irrelevant; for anything user-facing, it means you own the moderation layer entirely.
Reliability looks solid, performance is unreported
Uptime over the last 24 hours reads 99.986%, and over the last 30 minutes 99.989% — genuinely production-grade serving. But both the latency and throughput fields return null, so OpenRouter is publishing no speed data at all for this model yet. You cannot compare its tokens-per-second against anything without measuring it yourself.
The pricing cliff is undefined
There is no announced price for the period after the preview. A team that spends the week wiring Union Alpha into an agentic workflow is making a bet on a number nobody has published. Keep the integration behind a model-agnostic interface so swapping it out is a configuration change rather than a rewrite.
The exit plan writes itself
Given that the previous stealth model vanished from the catalogue entirely, the realistic planning assumption for Union Alpha is disappearance rather than continuity. Treat the preview as an evaluation window: learn what the model is good at, record it, and keep a paid model configured as the fallback that your agents fail over to when the free slot returns a 404.
Why Labs Keep Launching Models in Stealth
Two stealth launches in four weeks on the same two platforms is a pattern, not a coincidence. The incentives are straightforward once you look at what the lab gets.
Clean evaluation signal, unpolluted by brand
A named release is graded against expectations. A codename is graded on output alone, which gives a lab an unusually honest read on whether its model is actually competitive. Developers on OpenCode try Union Alpha because it is free and fast, not because they are loyal to or sceptical of whoever made it.
Enormous quantities of real agentic usage
A week of free access on a coding gateway generates far more genuine agent traffic than any internal evaluation harness — real repositories, real failures, real tool-calling sequences. Even under a no-training policy, aggregate telemetry about error rates, refusal patterns and context handling is valuable, and it is gathered before a single competitor sees the model.
A launch-day audience that no press release buys
The reach figures above make the point. A stealth drop on the right platform puts a model in front of hundreds of thousands of developers within hours, at zero marketing cost, with the mystery itself doing the distribution work. That is a strong argument for repeating the format, which is why it keeps happening.
The route into coding agents is now the route that matters
It is worth noting where these launches land. Not a chat app, not an API announcement — a terminal coding agent and a model router. The competitive front line for artificial intelligence model vendors has moved to whichever surface developers actually run their agents on, and stealth previews are how new entrants buy their way onto it.
Frequently Asked Questions About Union Alpha
Who made Union Alpha?
No lab has claimed it. OpenRouter states only that it is “developed and operated by a third-party model provider”. The routing metadata shows it serving on the Anthropic Messages API surface, which narrows the possibilities but does not identify the builder, since four Qwen models use that same surface on the same gateway.
How long will Union Alpha be free?
Both OpenCode and OpenRouter describe the preview as about a week from 16 September 2026, but neither has published an end date, and the OpenRouter listing carries no real expiry. Post-preview pricing has not been announced anywhere.
Is my code safe if I use it?
The two platforms disagree. OpenCode’s documentation claims a zero-retention policy; OpenRouter’s disclaimer says prompts may be retained but are not used for training. Both agree there is no training on your data. Given that neither company operates the model, treat the weaker claim as the operative one and avoid sending anything confidential.
How big is the context window really?
262,144 tokens, which is 256K exactly. OpenRouter’s post rounds it to “256K”. That is roughly 200,000 words, and it is a quarter of the one-million-token window the previous stealth model offered.
Can Union Alpha handle images?
Yes. The API reports input modalities of text and image, with text-only output. Screenshots, diagrams and UI mockups all work as input. Video is not supported, unlike its predecessor.
What happens to Union Alpha when the preview ends?
The honest answer is that nobody knows, but there is a precedent: the previous stealth model was removed from OpenRouter’s live model catalogue once its preview finished. Plan for the model to become unreachable, and keep a paid fallback configured.
Is it actually good?
There are no published benchmarks, no model card and no independent evaluation yet. Early testers describe it as competitive on coding tasks, which matches the “frontier-level” framing, but nothing verifiable exists at the time of writing. That is the trade the stealth format asks you to accept.
References and Further Reading
OpenRouter – Union Alpha model page
OpenRouter – Stealth API and Models
OpenRouter – Public models API
OpenCode – Zen gateway documentation
OpenRouter announcement post on X
OpenCode announcement post on X
CryptoBriefing – OpenCode launches Union Alpha for free use on OpenRouter
OfficeChai – Stealth model Ox Alpha free for a week
WinBuzzer – Stealth frontier model Ox Alpha appears free of charge
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.