At 14:42 UTC on September 16, 2026, a new entry appeared in OpenRouter’s model catalog with the id stealth/union-alpha, a 262,144-token context window, image input, tool calling, a price of $0, and a provider called Stealth. Ten minutes later OpenCode told its users it was free for a week. Twenty-seven minutes after that, OpenRouter’s CEO posted a chart with the model as a gold star near the top of the DeepSWE board and no number beside it. For about 33 hours, nobody said who made it. At 23:24 UTC on September 17 OpenRouter named it: Pareto, by Unbiased, a blended model that runs other labs’ models under one id, and the free window closed with the name.
This is the record of Union Alpha, resolved on the same URL the way our Ox Alpha record was in August. The reveal and the scored claims come first; the launch-day record follows as written, with its times.
On this page · 10 sectionsOpen
Union Alpha appeared in OpenRouter’s catalog as stealth/union-alpha at 14:42 UTC on September 16, 2026: 262,144-token context, 131,072-token max output, text and image input, tool calling and JSON output, priced at $0 for both input and output. The provider field reads Stealth; no lab, model card, weights or parameter count is attached.
OpenCode announced it at 14:52 UTC as free for the next week; the week lasted 33 hours. At 23:24 UTC on September 17 OpenRouter said the free stealth period had ended, the stealth listing lost its endpoints, and the named model went live at $2.50 per million input tokens, $0.25 cached, $7.50 output.
The numbers arrived with the name. Unbiased’s model card scores Pareto 26.9 against Fable 5.1, GPT-6 Astra and DeepSeek 4.1 Flash: DeepSWE 74 (a three-way tie at the top, Fable 67), Terminal-Bench 4.0 51 against 58 for Astra, MMMU-Pro 78 against 87, HLE no tools 49 against 55, ArXivMath 88 against 91. Vendor-run, no task costs, no composite; the launch-day chart claims stay unscored by anyone independent.
The data terms are narrower than the launch posts read. OpenRouter’s listing says prompts and completions may be retained by the provider but are not used for training, with everything else governed by the Stealth Model Terms. That is a no-training promise from an anonymous party, not a no-retention one.
The maker is Unbiased, the AI platform of Circuit & Chisel, a startup founded by two former Stripe engineers. Nobody fingerprinted it; OpenRouter credits detective work and demand for cutting the stealth period short, and the reveal came from the operator, as it did for Ox Alpha. The Manifold market’s 57 percent on Other was, for once, the right call.
The playbook held its record: Ox Alpha ran anonymous for six days, Union Alpha for 33 hours, and both free windows closed the moment the name was published. Pareto is not a new set of weights at all: Unbiased describes a blended model that runs several frontier and open models on each request and keeps the best answer, which is why no fingerprint matched a single lab.
§ 01The reveal: Pareto, by Unbiased
OpenRouter’s post: “Impressive detective work and demand from the community cut this stealth period short”. The stealth listing now reads “Union Alpha was a stealth model, revealed to be Pareto by Unbiased. OpenRouter routed requests to it and was not its developer, owner, or provider.” and “The free stealth period has ended”; its endpoints list is empty. The named model, unbiased/pareto, was created in OpenRouter’s catalog at 23:02:58 UTC, 22 minutes before the post.
| Item | Value |
|---|---|
| Model id | unbiased/pareto (stealth/union-alpha keeps its page, no endpoints) |
| Maker | Unbiased, the AI platform of Circuit & Chisel |
| What it is | A blended model: several frontier and open-source models run on each request, the best answer is kept; one model string, one bill |
| Price | $2.50 per million input tokens, $0.25 cached input, $7.50 output (OpenRouter and the Unbiased model card agree) |
| Context and output | 262,144 tokens in, 131,072 out; text and image input; tools and tool_choice |
| Provider | One, Unbiased; OpenRouter forwards every request to it |
| Version | The model card is titled Pareto 26.9 |
| Free window | 14:42 UTC September 16 to 23:24 UTC September 17, about 33 hours |
Unbiased’s own description matters for the identity question this page spent a day on: “Pareto is our blended AI model, available through the Unbiased AI platform.” Its homepage adds that “Pareto never switches models mid-conversation, so your prompt cache, and the savings, stay intact”, which is its argument for not being a router. OpenRouter’s listing calls it “a multimodal composite model built for research, coding, and agentic workflows”. Every guess under the launch posts named a lab. The answer was a company that sells a selection of labs.
Circuit & Chisel, per press coverage of its September 2025 seed round, was founded by former Stripe engineers Louis Amira and David Noel-Romas and raised $19.2 million led by Primary Venture Partners and ParaFi Capital, originally for an agent payments protocol; the company’s site says Unbiased and Pareto are built and maintained by that team. Those facts are reported, not read from a filing.
| Claim on September 16 | What the reveal says | Verdict |
|---|---|---|
| Free for a week (OpenCode) | Free window closed after 33 hours | Miss |
| About 73 percent on DeepSWE (Atallah’s chart) | Model card: 74 | Hit, vendor-scored |
| About 52 percent on Terminal-Bench v4.0 at about $1.60 a task (OpenCode’s chart) | Model card: 51; no task cost published | Score close; cost unverified |
| Outperforms GPT-5.6 Sol on Terminal-Bench 2.1 and SWE-Bench Verified | Neither benchmark on the model card | Unscored |
| Outperforms Opus 5 on Terminal-Bench 2.1 | Not on the model card | Unscored |
| Made by Kimi, MiniMax or Z.ai (replies, Manifold) | A blended model from Unbiased; Manifold’s Other at 57 percent | Other was right |
| No training on prompts, possible retention (listing) | Listing wording unchanged after the reveal | Stands |
| Benchmark | Pareto 26.9 | Fable 5.1 | GPT-6 Astra | DeepSeek 4.1 Flash |
|---|---|---|---|---|
| DeepSWE | 74 | 67 | 74 | 74 |
| Terminal-Bench 4.0 | 51 | 56 | 58 | 31 |
| MMMU-Pro | 78 | 81 | 87 | 77 |
| HLE (no tools) | 49 | 55 | 54 | 39 |
| ArXivMath | 88 | 72 | 91 | 28 |
The card is candid about its limits: “Measured task costs and a composite score have not been published for this release.” A blended model’s headline risk is cost per completed task, since several models may run per request, and that is the one number the card does not carry. The launch-day charts put a cost on the x axis; the reveal took it off. It stays on the watch list below.
What follows is the record as it stood on launch day, September 16, kept as written because its times are the evidence of how the day went.
§ 02What OpenRouter published
Everything in the table is from OpenRouter’s public model API and listing page, read on September 16 between 19:09 and 19:16 UTC.
| Item | Value |
|---|---|
| Model id | stealth/union-alpha |
| Listed | September 16, 2026, 14:42:03 UTC (catalog created timestamp) |
| Developer | Undisclosed; provider field reads Stealth |
| Context window | 262,144 tokens |
| Max output | 131,072 tokens |
| Inputs, outputs | Text and image in, text out |
| Parameters exposed | max_tokens, temperature, top_p, tools, tool_choice, response_format |
| Reasoning control | None exposed |
| Tokenizer | Listed as Other |
| Price | $0 per million tokens, input and output |
| Data terms | May be retained by the provider, not used for training |
OpenRouter’s own description of the model reads, in full: “Union Alpha is a multimodal model built for research, coding, and agentic workflows, while delivering frontier-level performance across a broad range of general-purpose tasks. Union Alpha is a stealth model. It is developed and operated by a third-party provider who has chosen to remain anonymous during this preview. OpenRouter routes requests to it and is not its developer, owner, or provider.” The listing’s own FAQ adds: “Union Alpha was released on September 16, 2026.”
Two things in the table are worth a second look. The context window is 262K, a quarter of Ox Alpha’s 1M, which is closer to the shape of a production model than a demonstration. And the listing exposes no reasoning parameter at all, where Ox Alpha exposed mandatory reasoning with effort settings. Either the model does not reason in a way the API surfaces, or the operator hid the knob for the preview.
§ 03What the launch posts claim
OpenCode’s announcement, at 14:52 UTC, was four lines: “Union Alpha (stealth model) is free for the next week”, then three bullets: no data training, built for agentic coding, supports images. Asked about limits five minutes later, the account answered “capacity so far is 5T a day but potentially can go up”. The post had passed 880,000 views by 19:13 UTC.
OpenRouter’s CEO Alex Atallah posted at 15:19 UTC: “Here’s the DeepSWE result for Union Alpha, a stealth model we just launched.” The result is an image: DeepSWE score against average cost per task, with the frontier models as labelled curves and Union Alpha as a star marked expected pricing in the top-right corner, where the cheap and strong models live. Reading the chart, the star sits near 73 percent, level with gemini-3.8-flash, gpt-6-astra and claude-opus-5, at a cost close to zero. No number is printed for it. One second later, in the same thread: “This free model also outperforms 5.6 Sol on Terminal-Bench 2.1 and SWE-Bench Verified. And outperforms Opus 5 on TBench 2.1”.
OpenCode’s chart, posted at 14:58 UTC and carrying Artificial Analysis branding, is a different benchmark: Terminal-Bench v4.0 score against cost per task on a log scale. Union Alpha is again a star, marked anticipated pricing, at roughly 52 percent and about $1.60 a task, sitting on the drawn Pareto line between GLM-5.3-Flash (about 33 percent, about $0.30) and GPT-6 Astra at high effort (about 55 percent, about $4). Again, no number printed.
| Claim | Source | Status |
|---|---|---|
| Frontier-level performance across general-purpose tasks | OpenRouter listing text | Operator description, no benchmark attached |
| About 73 percent on DeepSWE at expected pricing | Atallah’s chart image, 15:19 UTC | Read off a chart with no printed score; not on any public DeepSWE table |
| Outperforms GPT-5.6 Sol on Terminal-Bench 2.1 and SWE-Bench Verified | Atallah’s reply, 15:19 UTC | Text claim, no numbers; not listed on tbench.ai or swebench.com |
| Outperforms Opus 5 on Terminal-Bench 2.1 | Atallah’s reply, 15:19 UTC | Text claim, no numbers |
| About 52 percent on Terminal-Bench v4.0 at about $1.60 a task | OpenCode’s chart image, 14:58 UTC | Read off a chart; a different benchmark version from Atallah’s claim; not on Artificial Analysis’s public model pages |
So the honest summary is one sentence: every number about Union Alpha so far was drawn by the people routing it, priced at a cost the model does not yet have, and printed on no leaderboard. That is not an accusation. Stealth previews exist to gather exactly the usage that makes independent numbers possible, and Ox Alpha’s operator charts turned out to be roughly right. It is a statement about what kind of number a chart-with-a-star is.
The first replies are the usual mix. Under Atallah’s post: “looks benchmaxxed really”, “It thinks for way too long”, and a question about whether it is a Kimi model. Under OpenCode’s, by early evening: “It’s so slow at the moment. I’ll come back later”. Free models on launch day are slow; that reading tells you about demand, not the model.
§ 04Who made it
Answered on September 17 at 23:24 UTC (see the reveal above); the section below is the launch-day record. Nobody had shown it, and the tools that usually answer this quickly are blunted. When Ox Alpha appeared in August, a 95-of-95 tokenizer probe matched the GLM-5 vocabulary within three days, and Z.ai’s error envelopes sealed it. Union Alpha’s tokenizer is listed as Other, which forecloses the token-count comparison until someone does it by hand, and as of this writing no fingerprint study, error-string comparison or refusal-pattern read has been published.
What exists is speculation, and it should be labelled as such. Replies under the launch posts guess Kimi K3.1 and MiniMax M3.1, on timing. A Manifold market opened at 15:47 UTC asks the question directly; at 19:16 UTC, with nine bettors, it had Other at 57 percent, Z.AI at 12 percent, Mistral at 6 percent and Moonshot at 6 percent. A prediction market with nine participants is a mood, not a source, and it is reported here as one.
The name is a clue only in the sense that it is not one. OpenRouter’s stealth codenames have been abstract nouns and animals (Quasar, Horizon, Pony, Ox); Union maps onto no lab’s naming scheme. It invites the reading that this is a joint release, which is exactly the kind of reading a codename is chosen to invite.
What would resolve it: a tokenizer probe against the candidate vocabularies, an API error envelope that matches a known vendor’s format, a video or audio input that answers when it should not, or the operator saying so. OpenRouter’s precedent is that the operator says so within about a week.
§ 05The data question
The launch post says no data training. The listing says something more specific: “Prompts and completions may be retained by the provider but are not used for training; all other use is governed by the Stealth Model Terms.” Those two statements are compatible and not equivalent. The first is the second with the retention clause removed.
Read plainly, the promise is that an anonymous operator will not train on what you send, and may keep it. Alex Atallah wrote the general version himself in February 2026, about an earlier stealth model: “Note: it’s a stealth model, so has logging enabled”. Nothing about Union Alpha suggests that has changed, and the listing’s wording says it has not.
This matters more for a coding model than for a chat model, because the prompts are repositories. Ox Alpha’s window left an open question about what happened to the code sent during its anonymous week; Z.ai never addressed it after the reveal. The same question is open here from the first hour. Our own position is simple: we treat a stealth endpoint as an unnamed counterparty, we do not route CellCog agents to it, and we would not send it a repository we could not publish.
§ 06The stealth playbook, and its record
OpenRouter has run this play before, and the record is short enough to hold in a table.
| Model | Appeared | Revealed as | Days anonymous |
|---|---|---|---|
| Quasar Alpha | April 3, 2025 | An early GPT-4.1 test, per OpenRouter, April 14, 2025 | 11 |
| Horizon Alpha | July 30, 2025 | An early GPT-5 checkpoint, per OpenRouter, August 7, 2025 | 8 |
| Pony Alpha | February 6, 2026 | Z.ai GLM-5, per OpenRouter, February 11, 2026 | 5 |
| Ox Alpha | August 20, 2026 | Z.ai GLM-5.3-Flash, August 26, 2026 | 6 |
| Union Alpha | September 16, 2026 | Open | Counting |
Quasar Alpha was the first, announced by Atallah on April 3, 2025 as “our first-ever stealth model”, and named by OpenRouter on April 14, 2025, GPT-4.1’s launch day, as an early test of that model. Horizon Alpha ran eight days before OpenRouter named it a GPT-5 checkpoint; Pony Alpha ran five before it became GLM-5. Ox Alpha appeared on August 20, 2026, drew roughly 221,000 users and 16 trillion tokens on OpenCode in three days, and was confirmed as GLM-5.3-Flash on August 26, at which point the free window closed and the paid route opened the same day. The pattern is the point: the free week is the launch campaign, the users burning the free tokens are the launch audience, and the reveal ends the price.
§ 07What we are watching for
- A task cost. Unbiased has not published measured cost per completed task for Pareto 26.9. For a model that runs several models per request, that figure decides whether $2.50 / $7.50 is cheap; the first one, from Unbiased or an outside run, goes in the scores table.
- An independent number. A DeepSWE or Terminal-Bench run by someone other than Unbiased. The launch-day chart claims about Sol and Opus 5 on Terminal-Bench 2.1 and SWE-Bench Verified stay unscored until then.
- Which models are inside. Unbiased names the comparison set, not the component models Pareto blends. A disclosure changes the data question: prompts sent to Pareto reach whichever labs it runs.
- What happened to the stealth week’s prompts. The listing still says retained, not trained on; Unbiased’s data policy governs the named model. As with Ox Alpha, the anonymous window’s data question stays open until the operator addresses it.
§ 08The tracker
| Question | Status |
|---|---|
| Who made it | Unbiased (Circuit & Chisel); Pareto, a blended model; revealed by OpenRouter September 17, 23:24 UTC |
| Is it free | No; window closed after 33 hours; $2.50 in, $0.25 cached, $7.50 out per million tokens |
| Frontier performance | Vendor model card: five scores against Fable 5.1, Astra and DeepSeek 4.1 Flash; no task costs, no independent run |
| Data terms | Stealth listing wording unchanged; Unbiased’s own data policy now governs the named model |
| Reveal | Done |
§ 09Update log
- October 1, 2026, 19:10 UTC. A second version. OpenRouter listed unbiased/pareto-26.10-preview at 14:07 UTC: Pareto 26.10 Preview, with a 1,048,576-token context (four times the 262,144 of Pareto 26.9), a 131,072-token max output, and text and image input, at $0.80 per million input tokens, $0.03 cached and $3.20 output, against $2.50 / $0.25 / $7.50 for Pareto 26.9. The listing says only that it is a preview; scores go in the table when Unbiased or an outside run publishes them.
- September 18, 2026, 01:30 UTC. Resolution. OpenRouter revealed Union Alpha as Pareto, by Unbiased (Circuit & Chisel), at 23:24 UTC on September 17, 33 hours after the listing; the free window closed, the named model listed at $2.50 / $0.25 / $7.50 per million tokens, and Unbiased published a model card with five vendor-run scores. Reveal section, scored-claims table and model-card table added; title, glance answers, FAQ and tracker rewritten; the launch-day record kept below with its times.
- September 16, 2026, 19:16 UTC. Page opened, four and a half hours after the catalog entry. Facts from the OpenRouter API and listing; posts read on X with times taken from their ids; benchmark charts read off the images, no scores printed; Manifold read as sentiment.
§ 10Sources
OpenRouter, Union Alpha listing and the public model API (/api/v1/models and /api/v1/models/stealth/union-alpha/endpoints), read September 16, 2026, 19:09 UTC (created timestamp 1789569723). OpenCode, post on X, 2026-09-16T14:52:40Z, and replies at 14:57:00Z and 14:58:28Z. Alex Atallah, post on X, 2026-09-16T15:19:12Z, and reply at 15:19:13Z; Quasar Alpha post, April 3, 2025; stealth logging note, February 8, 2026. Manifold Markets, Who is behind Union Alpha, created 2026-09-16T15:47:24Z, read 19:16 UTC. Leaderboard absences checked at 19:16 UTC on artificialanalysis.ai/models, tbench.ai/leaderboard and swebench.com/verified.html. OpenRouter blog, Quasar Alpha and Optimus Alpha reveal, April 14, 2025, and the Pony Alpha and Horizon Alpha model pages, which carry each model’s release date and its revealed identity. Ox Alpha facts from our record, sourced there. Reveal (September 17-18): OpenRouter, reveal post on X, 2026-09-17T23:24:41Z; OpenRouter, Pareto listing and /api/v1/models/unbiased/pareto/endpoints (created 1789686178), read September 18, 01:20 UTC; the Union Alpha listing after the reveal, read 01:15 UTC; Unbiased, Pareto 26.9 model card and homepage, read 01:21 UTC; Circuit & Chisel, company site; Fortune, Stripe crypto alum raise $19.2 million, September 22, 2025 (founders and round, press-reported).
Q1How do I use Union Alpha?
The stealth id stealth/union-alpha has no endpoints since September 17. Use unbiased/pareto on OpenRouter in any OpenAI-compatible client (one provider, Unbiased; tools and tool_choice supported), or the id pareto on Unbiased’s own platform. Same 262,144-token context and 131,072-token max output as the stealth listing.
Q2Is Union Alpha free?
Not since 23:24 UTC on September 17, 2026, when OpenRouter announced the reveal and said the free stealth period had ended. Pareto is $2.50 per million input tokens, $0.25 for cached input, and $7.50 per million output tokens, the same figures on Unbiased’s model card and OpenRouter’s listing. The free window ran 33 hours, against the week OpenCode announced.
Q3Does Union Alpha train on my prompts?
OpenRouter’s listing says prompts and completions may be retained by the provider but are not used for training. So: no training, possible retention, terms set by an operator that has not named itself. Alex Atallah noted in February 2026 that stealth models have logging enabled. Send it what you would send an unnamed counterparty.
Q4How good is it really?
Unbiased’s model card publishes five scores for Pareto 26.9: DeepSWE 74, Terminal-Bench 4.0 51, MMMU-Pro 78, HLE (no tools) 49, ArXivMath 88, beside Fable 5.1, GPT-6 Astra and DeepSeek 4.1 Flash. The card says measured task costs and a composite score have not been published. The DeepSWE score matches the roughly 73 percent read off Atallah’s launch-day chart; the Terminal-Bench 4.0 score is close to OpenCode’s chart, without the $1.60 cost figure. Still no independent run.
Q5Which lab is behind Union Alpha?
Not a lab. Unbiased, by Circuit & Chisel, whose Pareto runs several frontier and open-source models on each request and keeps the best answer; Unbiased says it is not a router because it never switches models mid-conversation. The Kimi and MiniMax guesses were wrong, and so was the framing of the question: there was no single lab to find.
Q6What happened with the last stealth model?
Ox Alpha appeared on OpenRouter on August 20, 2026, passed 221,000 users on OpenCode in three days, and was confirmed on August 26 as Z.ai’s GLM-5.3-Flash, with MIT weights the same day. The tokenizer fingerprint called the lab before the reveal. Our record of it is at /blog/what-is-ox-alpha/.