Union Alpha vs DeepSeek V4 Flash (Free): What "Free" Actually Means on Each One

There is a specific failure mode that eats a day of engineering time, and it looks like this: you find a model listed at $0, wire it into a script, and the first real call comes back with a provider error rather than a clean 404. The listing said "free" and nothing was behind it. That is the trap Union Alpha and DeepSeek V4 Flash (Free) sit on opposite sides of — one is free because nobody has started charging yet, and the other is free because the price is genuinely zero and somebody is still paying to serve you. Both are on OrcaRouter's free tier today, and they will behave very differently the moment the free part ends.

Two kinds of free, and only one of them is a price

A free preview is a marketing window. The operator eats the compute to collect usage data, telemetry, and mindshare, and the price is a temporary state that will be replaced. A free tier is a product decision: the vendor has concluded that some slice of traffic is worth serving at zero to pull developers into a paid path, and the zero is meant to persist.

Those two things look identical on a pricing page. They behave nothing alike in production.

Union Alpha: free because the meter has not started

Union Alpha appeared on third-party catalogues on 16 September 2026 and is served on OrcaRouter's free tier. Its operator is anonymous — the provider field reads "Stealth", no lab has claimed it, and there is no model card, no licence, no parameter count, and no published benchmark.

What you know for certain is the endpoint:

• Context window — 262,144 tokens, with 131,072 tokens of maximum output.

• Inputs — text and images; output text only. No video, no audio.

• API surface — tools, tool_choice, response_format, temperature, top_p, max_tokens. JSON output yes; strict JSON-schema enforcement no.

• Reasoning controls — none. There is no effort parameter to turn up or down.

• Price — $0 per million input and output tokens, rate-limited, HTTP 429 past your plan cap. Usage is never charged to your balance.

• The clock — described as a one-week free window, with no post-preview pricing announced.

One independent test exists: the SMF Clearinghouse Official A suite, 157 tests, reasoning off, scored 136/157 (86.6%) with zero errors, ranking tenth of twenty-six. Strong on reasoning (30/30) and tools (2/2), decent on coding (26/30) and math (24/30), visibly weak on writing (2/5). That is the entire public performance record.

DeepSeek V4 Flash (Free): free because zero is the price

DeepSeek V4 Flash is not a stealth model. It is a 284B-parameter mixture-of-experts model with roughly 13B parameters active per token, released under the MIT licence, with a 1M-token context window and up to 384K tokens of output. Its free variant on our free tier is priced at $0 per request, rate-limited, with HTTP 429 past the cap and a paid sibling available on the same key when you outgrow it.

The distinction that matters is who is paying. On DeepSeek V4 Flash (Free), nobody is running a countdown. DeepSeek V4 Flash has a published paid rate — a published paid rate for the same model on the same key, and the free tier is a deliberate on-ramp to it. The model has a licence, a lineage, and a documented architecture. If the free tier vanished tomorrow, you would not be stuck; you would be paying a known rate for a model you already understood.

That is not a small thing. It converts "the free thing disappeared" from an outage into a line item.

The risks that come with each free tier

Retention. Free endpoints are where the least favourable data terms live. On Union Alpha, the operator has stated that prompts and completions may be retained by the provider but are not used for training — and one distribution channel has separately promised zero retention, which is a contradiction rather than a clarification. Treat "not used for training" and "not stored" as different claims, because they are. If your input contains anything you would not paste into a stranger's form, this is not the endpoint for it.

Rate limits. Both are rate-limited and both return 429. Free previews and free tiers share a pool, get deprioritised under load, and have caps that flex with demand rather than being published. Any production-ish loop needs exponential backoff and a fallback path regardless of which one you pick.

Single point of failure. Union Alpha runs on a single endpoint with no fallback provider behind it. There is no second route if it degrades, which is precisely the situation automatic failover exists to paper over.

The retirement question, which is the real comparison

Stealth previews have a documented pattern of disappearing. The previous anonymous entry in this series, Ox Alpha, was revealed six days after appearing as Zhipu's GLM-5.3-Flash and was subsequently pulled from catalogue. Union Alpha's free window has been described as roughly a week. Nobody has said what happens on day eight.

DeepSeek V4 Flash (Free) has a different endpoint-of-life risk, and it is subtler: obsolescence rather than disappearance. DeepSeek V4.1 Flash, released 10 September 2026, is absorbing legacy V4 Flash and V4 Pro requests at Flash pricing. A free tier on a model the vendor is consolidating away from is not a crisis — the paid sibling still works and the rates are published — but it is a signpost worth reading before you build a long-lived integration on the free variant specifically.

There is also a broader caution about free listings generally. Free endpoints for popular models are frequently observed in a broken state: the badge says "Free" while the endpoint list behind it is empty, so a call fails with a provider error instead of a clean model-not-found. If you are evaluating any $0 listing, call it once before you design around it.

Running both without a second integration

The practical advantage of putting both on one key is that the free-tier question stops being architectural. OrcaRouter passes provider list price through at 0% markup, so when a vendor moves a rate — as DeepSeek did across the V4 line on 16 August — the change is live on our side the same day. Point your evaluation at the free tier, keep the paid sibling on the same key, and add an automatic failover rule so a 429 or a vanished endpoint falls through to a model that is still answering instead of surfacing as a stack trace.

That is the whole shape of it: use the free thing as a free thing, and never let it be the only thing on the path.

Which free model should you pick?

Pick Union Alpha if you want to find out what the mystery model can do. Its vision input and 256K window are genuinely useful for screenshot and document work, the Official A result suggests it is competent, and the cost of being wrong is a week of evaluation that you were going to spend anyway. Keep it in a sandbox, keep it off sensitive input, and assume the endpoint can vanish.

Pick DeepSeek V4 Flash (Free) if you want a free model you can build a habit around. The 1M-token context, 384K output ceiling, MIT licence, and a paid rate you can step onto without changing anything are what make it the durable choice of the two. You are trading the novelty of an anonymous frontier claim for the boring reliability of a model whose owner has a name.

The version of this decision that ages well is the one where you never have to make it in a hurry. Put both behind one key today, watch what the free preview does when its window closes, and let the failover rule hold the line in the meantime.

The durable half of this comparison is documented: DeepSeek V4 Flash (Free) model page lists the free tier's $0 rate, its 429 behaviour past the cap, and the paid sibling on the same key.