

Union Alpha and the quiet price of a free frontier model
A stealth model called Union Alpha is free until about 22 September. Before you paste your codebase into it, learn what zero data retention AI really guarantees, and what it politely leaves out.
By Katie Delaney / 2026-09-16 / 11 min read

Union Alpha: free for a week, with no name on the door
The fox does not sprint at every rustle in the brush. It watches the trail, checks the wind, and only then moves. That patient prowl is the right posture this week, because a stealth model called Union Alpha has appeared on OpenRouter, free for roughly seven days, carrying frontier-level claims and no named maker. Before you paste a single secret into it, learn what zero data retention AI actually guarantees, and what it conveniently leaves out.
The listing describes Union Alpha as a multimodal model built for research, coding and agentic workflows, with frontier-level general performance, according to OpenRouter's Union Alpha listing. The headline numbers are a 262,144-token context window, text and image input, tool calling and reasoning support, all at zero price during the preview, as recorded in the same listing and the models registry entry. OpenCode itself announced the launch on X as a stealth model that is free for the next week, built for agentic coding, with image support, in OpenCode's launch post. That combination, a big context plus agent tooling plus free, is exactly why the hunting party formed within hours.
Union Alpha (stealth model) is free for the next week - no data training - built for agentic coding - supports images let's see what you can do
OpenCode is an open source AI coding agent that runs in the terminal, on the desktop or inside an editor, as its own documentation explains, so a model tuned for agentic loops slots straight into its daily workflow. Early user reports mention detailed image-grounded work and long coding sessions, but those are pub stories until somebody publishes a rerun, and this article treats them that way.
Shoppers hunting the best free ai model for coding should read this launch as a timed trial, not a gift. The price is zero, the meter starts now, and the quiet quarry is your usage data, your prompts and your patience. That is why zero data retention AI headlines every serious review of this launch.
What you get for zero pounds
A context window of 262,144 tokens swallows whole repositories in one pass, image input lets the model read screenshots and diagrams beside the code, and tool calling lets it act inside an agent harness rather than merely chat. The registry entry confirms attachment handling, reasoning and tool use, with closed weights, per the models registry entry. Put plainly, this is a borrowed brain for a week: sharp scent, strong legs, and a den you do not own. That is a fine deal for experiments, and a poor foundation for anything you cannot afford to move. Free, fast and masked is a thrilling combination, and it is precisely when zero data retention AI discipline matters most.
| Item | Value |
|---|---|
| Union Alpha | 262144 |
| Ox Alpha | 1048576 |
| Qwen3.8-27B | 262144 |
Union Alpha meets zero data retention AI: read the wording
Here is where the careful calibration starts, because three wordings are circulating and they do not say the same thing. OpenRouter's listing says prompts and completions may be kept by the provider but are never used for training. CryptoBriefing describes a zero-retention week. OpenCode's launch post says there is no data training. Those statements overlap, but no training is not the same as no keeping, and a prudent posture treats kept prompts as published prompts. This is the whole of zero data retention AI in one sentence: the guarantee covers the model's future weights, not the provider's present logs.
Buyers already ask these questions in search whenever trust runs thin, under queries such as open ai zero data retention and perplexity ai data retention, alongside policy-shaped searches like open ai data retention policy. The pattern is steady scrutiny: shoppers compare retention wordings before they compare benchmarks. The shortest honest phrase in that cluster is zero data retention AI, and it earns its place in this article because it names the exact anxiety this launch triggers. For regulated teams, that anxiety is not a mood, it is a governance ticket, and our AI consultancy exists for weeks like this one.
Open the model listing and find the sentence about kept prompts. If it permits keeping, assume keeping, whatever the launch post says.
Point the agent at a clean checkout with no secrets, no customer data and no private keys. Experiments deserve a sandbox, not your vault.
Remove hostnames, tokens, staff names and client references before the first prompt. Redaction is den discipline for grown-ups.
Review generated patches before they touch shared branches. A borrowed brain can still write a memorable mistake.
Note what you sent, what it cost in time, and what you would pay to keep. Free trials end, and measured notes beat fond memories.
One more caution from the practitioner thicket: a router sits between you and the model, and a lively thread this week asks how anyone verifies which model actually answered, as the router-trust debate records. HTTPS proves you reached the router, not which weights ran. That does not make routers villains, it makes blind trust expensive. Test steadily, keep receipts, and let the measured middle guide you.

Union Alpha benchmarks are vendor postcards until somebody reruns them
The loudest numbers come from the launch team itself. OpenCode's founder shared a DeepSWE result for Union Alpha with an invitation to try it in any harness, and added that the free model beats a named frontier rival on Terminal-Bench 2.1 and SWE-bench Verified, plus a second frontier model on Terminal-Bench, in the founder's post. A community thread goes further, claiming wins on DeepSWE Pro against three named models, in the community thread.
These are postcards, not maps. No independent rerun has been published at the time of writing, so every figure below sits in the evidence table as vendor-reported or community-claimed, never as established fact. Hold them to the same zero data retention AI standard as the privacy wording: verify first, then trust.
What the tests actually measure is public and solid. Terminal-Bench scores agents on real terminal tasks and is run as a Stanford collaboration with a public leaderboard, per Terminal-Bench. SWE-bench Verified is a human-filtered set of 500 real software issues scored by percentage resolved, per the SWE-bench leaderboards and OpenAI's Verified report. DeepSWE-style agentic coding runs use harnesses like the one behind the published Ox Alpha result of 58.4 percent, documented step by step in the Ox Alpha bench log. Knowing the machinery matters, because it tells you what a rerun must show before the applause starts.
| Claim | Source | Status |
|---|---|---|
| Beats named rivals on DeepSWE Pro | Community thread | Unverified, single anecdote |
| Tops rivals on Terminal-Bench 2.1 and SWE-bench Verified | Launch team post | Vendor-reported, no rerun yet |
| Matches GPT-6 Astra on coding benchmarks | Launch brief | No source found, unverified |
| Detailed SVG and long code sessions | Early user reports | Unverified, illustrative only |
| Free for about a week from 16 September | Launch post and press | Confirmed by two sources |
| 262K context, image input, tool calling | Model listing and registry | Confirmed by two sources |
- Beats named rivals on DeepSWE ProCommunity threadUnverified, single anecdote
- Tops rivals on Terminal-Bench 2.1 and SWE-bench VerifiedLaunch team postVendor-reported, no rerun yet
- Matches GPT-6 Astra on coding benchmarksLaunch briefNo source found, unverified
- Detailed SVG and long code sessionsEarly user reportsUnverified, illustrative only
- Free for about a week from 16 SeptemberLaunch post and pressConfirmed by two sources
- 262K context, image input, tool callingModel listing and registryConfirmed by two sources
| Item | Value |
|---|---|
| 2026-03-11 | Hunter Alpha |
| 2026-03-12 | Healer Alpha |
| 2026-08-20 | Ox Alpha |
| 2026-09-16 | Union Alpha |
The forensics playbook already exists, because this film has played before. When Hunter Alpha appeared, analysts matched training cutoffs, compared reasoning style and traced a sibling model to an earlier confirmed launch, while dissenters pointed at token behaviour that did not fit, as Reuters, the Independent and the Awesome Agents analysis variously report. Expect the same measured middle now: tokenizer probes, error-string dialects and chain-of-thought tics, weighed against the precedent that the same masked account previously served a different lab's model. Trail talk is fun, but hard evidence wins.
So who built Union Alpha
Six suspects stand in the moonlit lineup, and each has a thread of evidence, however thin. Mistral has the name theory, since Union reads as a nod to the European Union, plus a community-noted match on the 262K context size and a conveniently timed summit appearance a week out, per the Mistral thread. Moonshot AI has the Kimi rumour mill. DeepSeek has the standing habit of being blamed for every masked launch. Zhipu has the precedent of a previously confirmed stealth outing. OpenAI has the motive of quietly testing coding strength, while Grok and Claude are the long shots that keep the poll honest.
None of them has claimed it, and OpenRouter says the provider chose silence, per the listing. That silence is itself a datum: somebody is buying a week of real-world traffic before a reveal.
My guess, stated as opinion and worth exactly what you paid for it: Mistral. The European nod in the name, the summit timing that fits the usual reveal-a-week-later rhythm, and a context size the community already associates with their small models make the vulpine case. DeepSeek runs second, mostly on base rates rather than fingerprints. If the mask slips and it is Zhipu or Moonshot instead, the lesson still holds, because the winning move was never picking the lab, it was testing the model while the price was zero. Whichever lab claims it, the search for the best coding ai model free simply moves to the next free window.
If the mask slips in a week, will you be surprised, or will you have called it?
That is the question to carry into the free week, and the game is open to everyone. Run the five checks, keep your secrets out of the prompt box, note which benchmark claims survive contact with your own repository, and post your guess with receipts. The fox admires a bold call, but it respects a documented one. For teams that want this kind of silent switchboard watched properly, our SEO and GEO work and cybersecurity marketing do exactly that, and the hedgerow chatter on the news desk keeps the story updated. Now tell us: who do you think built Union Alpha, and what is your evidence?
Frequently asked questions
What is Union Alpha?
Union Alpha is a stealth AI coding model released on 16 September 2026 through OpenRouter in collaboration with OpenCode. It handles text and image input, offers a 262,144-token context window with tool calling for agentic workflows, and is free during a preview of roughly one week.
How long is Union Alpha free?
About a week from launch, so until roughly 22 September 2026, according to the launch post and press coverage. Pricing after the preview has not been announced, so treat the window as a timed trial and finish your evaluation inside it. That timing matters for zero data retention AI planning: decide what may be sent before the week runs out.
Who built Union Alpha?
Nobody has claimed it. OpenRouter says an anonymous third-party provider operates it. Community guesses include Mistral, OpenAI, DeepSeek, Moonshot AI, Zhipu and Claude. Our opinion leans Mistral on the name, timing and context-size threads, but that is a guess, not a finding. Cast your vote in the poll below.
Does zero data retention AI mean my prompts are deleted?
Not necessarily. Zero data retention AI usually means prompts are not used for training, while the provider may still keep logs. Union Alpha's listing permits the provider to keep prompts and completions. Assume anything you send is stored, and keep secrets out of the prompt box.
How are DeepSWE and Terminal-Bench scored?
Both run agents against real tasks and count successes. Terminal-Bench tests terminal mastery with a public leaderboard run as a Stanford collaboration. SWE-bench Verified scores the percentage of 500 human-filtered software issues resolved. Union Alpha's claimed wins have no published independent rerun yet. Until one lands, give every figure the full zero data retention AI treatment: verify first, then trust.
Can I use Union Alpha inside OpenCode today?
Yes. The launch team says it works in any harness under the model id stealth/union-alpha, and OpenCode's documentation covers connecting providers and starting sessions. Use a clean checkout with no secrets for your first runs, and review patches before they reach shared branches.
The Agents API makes autonomy a governance decision
Read the field notefolkfoxHow to build a meaningful AI search visibility dashboard
Read the field notefolkfoxGenerative engine optimisation services: what AI Overviews can actually cite
Read the field notefolkfoxOpenAI for Finance: the sharp test for AI agents
Read the field noteWho built Union Alpha? Cast your vote
Six suspects, one mask. Tap your guess and the shared tally updates instantly. One vote per browser per day, counted anonymously.
Fetching the tally…
No accounts and no tracking beyond a daily anonymous lock that stops repeat votes.
Testing Union Alpha this week?
Bring our zero data retention ai checklist to the trial. We help teams evaluate AI coding models safely, then ship the ones that earn it.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

