Free and “Zero Retention”? Picking a Model in OpenCode for Sensitive Work Taught Me to Ask Who's Making the Promise

Contents

1. The top-scoring model was the first one I dropped
2. The comparison: cheap ones have loose terms, clean ones want your card
3. Digging in: what “zero retention” actually guarantees
4. How I’m handling the uncertainty
5. A few other things worth noting
6. One thing to take with you

Sincere words are not beautiful; beautiful words are not sincere.
—— Laozi, Tao Te Ching, ch. 81 (Spring and Autumn period); translation mine

On October 17, 2026, I wanted a model inside OpenCode, the AI coding tool, that I could trust with sensitive material: no retention, no training on my data, good enough to use, and cheap. I ended up with DeepSeek V4.1 Flash Free, which OpenCode offers at no cost. Its docs say the provider “follows a zero-retention policy and does not use your data for model training,” and the specs hold up (a 200,000-token context, image input, reasoning). The catch is that this is the provider’s own statement, the docs never say who the provider is, and I couldn’t find it anywhere on my machine either. So I emailed OpenCode the same day. Until they answer, I use it, but the truly confidential stuff stays out.

The top-scoring model was the first one I dropped

I started the obvious way: open the free-model list, set the highest scorer as default. It was labeled Contributor. Shortly after, I worked out what that label meant. The price of using it is that your conversations get used to train someone else’s model. It’s free because you are the raw material.

That flipped my whole order of operations. I had been checking capability, then price, and skimming the terms last. Now the terms go first, and only what survives gets compared on score and cost. If you need confidentiality, “free” and “safe for sensitive work” have nothing to do with each other.

The comparison: cheap ones have loose terms, clean ones want your card

Next I lined up every route I could plug into OpenCode. Prices are USD per million tokens, input/output, checked on 2026-10-17:

Option	Price (in/out)	What happens to your data	Verdict
Top-scoring free Contributor model	0	Used for training	Out
OpenCode’s DeepSeek V4.1 Flash Free	0	Docs say zero retention, no training; host unnamed	Picked
OpenCode paid DeepSeek V4.1 Flash	0.30 / 1.20	Paid terms	Pricier than going direct; skipped
DeepSeek’s own API (off-peak)	0.15 / 0.60	Policy says data may be used to improve models, no clean opt-out	Out
DeepSeek V4.1 Flash on DeepInfra	0.20 / 0.60	Zero retention by default per its own policy	Paid backup

The irony is plain. The cheapest first-party API has the loosest terms. DeepInfra, with the cleanest terms, won’t serve you until you add a card or prepay. And the free model sits right where “costs nothing” meets “docs say no retention,” so it looks like the obvious answer.

“Looks like” is where the trouble starts.

Digging in: what “zero retention” actually guarantees

Here’s the sentence from OpenCode’s docs:

DeepSeek V4.1 Flash Free is free on OpenCode for a limited time. Its provider follows a zero-retention policy and does not use your data for model training.

Pull it apart and it rests on three legs, and each one wobbles a little.

Leg one: who is speaking. It says its provider follows zero retention, not OpenCode’s own servers. Your data passes through OpenCode and then goes to a third party to run. Which third party, the docs don’t say. I went through OpenCode’s local session database, logs and model catalog, and every record lists the upstream simply as “opencode.” There’s no field naming the real host. You can’t read the privacy policy of a company whose name you don’t know.

Leg two: is it a commitment or boilerplate. When I rechecked the docs, I noticed that another free model on the same page, GLM-5.3-Flash Free, carries the exact same sentence, word for word. That doesn’t make it false. But it reads more like a standard line for the free tier than a model-specific statement backed by a contract. Whether there’s a data processing agreement, a technical setup where the host simply can’t store anything, or an outside audit, the docs don’t say.

Leg three: how long it lasts. “For a limited time,” with no announced end date. And with the host anonymous, if the provider ever changed, I’d have no way to notice from my side.

So the claim holds only under narrow conditions: it’s a self-reported statement, you trust OpenCode to vet its upstream, and nothing changes while you’re using it. I can’t verify any of the three.

Why pick it anyway? Because the other two routes have problems I can verify. The first-party API says outright that it may use your data. The Contributor model is even blunter. One “says it won’t, can’t be checked” against two “says it will” — the first is at least uncertainty pointed in the right direction. It’s like renting a room where the landlord promises verbally not to come in. You can’t install a camera to check, but it beats a lease that says the landlord may enter anytime. You still don’t leave your valuables on the desk.

How I’m handling the uncertainty

Once that was clear, I did three things.

First, I drew a line around sensitive data. Everyday coding, organizing public material, editing posts: the free model is fine. Truly confidential files, personal information, contract terms: never pasted in. The model doesn’t hold that line. I do.

Second, I asked directly. That afternoon I emailed OpenCode’s support address with four questions: which upstream provider serves this model; whether zero retention rests on a contract, a technical mechanism or an audit; whether “no training” also rules out anonymized or aggregate uses, and whether OpenCode’s own gateway keeps prompts even briefly; and whether the free period has an end date. As of writing, no reply yet.

Third, I set up a paid fallback without paying yet. OpenCode has built-in DeepInfra support, so it’s just pasting a key, no config edits. I gave myself three triggers, any one of which means loading $10 and setting a spending cap: the free quota runs out, I need more than 200,000 tokens of context, or OpenCode’s answer can’t point to a real basis.

Halfway through I did ask myself whether I should just top up now to be safe. No. Paying today doesn’t buy more certainty. It swaps one provider’s uncertainty for another’s and adds a bill.

A few other things worth noting

• The free model only works inside OpenCode. Calling its endpoint directly from outside returns a free-tier access error. You can’t wire it into other tools.
• Router-site policy tables are secondhand. Some aggregators list which hosts are zero-retention, and a few entries contradict what the hosts themselves say. If the host’s own page doesn’t say it, I don’t count it.
• A host’s policy may not cover models it relays. Calling another big lab’s model through DeepInfra doesn’t necessarily come with DeepInfra’s zero-retention promise.
• Prices move daily. DeepInfra’s catalog went from 180 to 181 models between morning and evening on the same day. Every number here is a 2026-10-17 snapshot.

One thing to take with you

When “free” and a guarantee show up in the same sentence, find out who’s doing the guaranteeing. What this taught me wasn’t which model is strongest. It was to ask who stands behind a nice-sounding promise, whether they have a name, and whether you can hold them to it.

Here’s a small exercise for today. Open a free service you use a lot, whether it’s email, cloud storage, a translator or an AI tool, and find the line where it says what it won’t do with your data. Ask three things: who is saying this? Is it in a contract or on a marketing page? If they changed the wording, would you even know? Whichever question you can’t answer is where you should draw your own line.

Raw Markdown