The True Cost of Free: Who's Training on Your Prompts?

Free AI models sound great — until you realise your code, conversations, and workflows are becoming someone else's training data. Here's what every provider actually does with your data.

2026-04-21 · 6 min read

That "free" AI model? It's not free at all. You're paying with your most valuable asset: your data. Every prompt, every line of code, every conversation — it's all being harvested to train someone else's commercial models. I realised this too late, after watching my agent workflows unwittingly feed NVIDIA's next generation.

Let me pull back the curtain on what providers actually do with your data when you think you're getting something for nothing.

Free Isn't Free

There's a disclaimer on every free model page on OpenRouter that most people never read:

"For the free endpoint, all prompts and output are logged to improve the provider's model and its product and services. Please do not upload any personal, confidential, or otherwise sensitive information."

Let me translate: your prompts are training data. That's the deal. You get free inference; they get free training data. It's a fair exchange — if you know about it.

The problem is, most people don't. They see "free" and start running proprietary code, business logic, and agent workflows through these endpoints without understanding what happens next. Your code patterns, your architecture decisions, your debugging strategies — they all feed into the next version of someone's commercial product.

The Provider Scorecard

I spent an afternoon cataloguing every major provider on OpenRouter and Ollama to see who trains on your data and who doesn't. Here's what I found.

OpenRouter: Three Categories of Provider

Safe (No training, ZDR — Zero Data Retention):

• Google Vertex (41 models)
• Amazon Bedrock (23 models) — also has content moderation
• Anthropic (10 models) — also has content moderation
• Z.ai (13 models, 1 free)
• DeepInfra (73 models)
• NovitaAI (70 models)
• Together (27 models)
• Groq (8 models)
• SiliconFlow (34 models)
• Alibaba Cloud (39 models)

Grey Area (No training badge, but no ZDR either):

• OpenAI (60 models) — "No training" but no ZDR badge. Moderated.
• DeepSeek (1 model) — No training badge, no ZDR. China-based.
• StepFun — No training, no ZDR.
• Moonshot AI — No training, has ZDR.

Trains on your data:

• NVIDIA (5 models, 5 free) — The ONLY major provider on OpenRouter without a "No training" badge. All 5 of their models are free-tier only. Your data is the product.

Let me spell that out: Every free model on OpenRouter that routes through NVIDIA trains on your prompts. Nemotron 3 Super, Nemotron 3 Nano, Nemotron Nano VL — all of them. If you're using these models for coding, your code is becoming NVIDIA's training data.

The Free Model Trap

Here's the catch-22: OpenRouter lets you set a privacy preference to "deny routing to providers that train on your data." But if you enable that setting and you're on the free tier, almost no models work. Because almost all free models route through providers that train.

The free models exist to generate training data. That's the business model. NVIDIA isn't running 120B parameter models for free out of the goodness of their hearts — they're doing it because your prompts are worth more than the compute costs.

Ollama: A Different Story (With Caveats)

Ollama's approach is fundamentally different. Their privacy policy states:

"Ollama runs locally. We don't see your prompts or data when you run locally... Your data stays on your machine."

For locally-run models, this is true by design — the compute happens on your hardware, no data leaves the machine. But Ollama now offers cloud models (like glm-5.1:cloud, gemma4:e4b-cloud), and here's where it gets nuanced:

1. Ollama's own claim: "Prompt or response data is never logged or trained on."
2. Their providers: Ollama uses NVIDIA Cloud Providers (NCPs) and claims to require "no logging, no training, and zero data retention policies" from partners.
3. The gap: Ollama hosts primarily in the US, with routing to Europe and Singapore for capacity. They partner with NVIDIA — the same NVIDIA that trains on data through OpenRouter. The difference is that Ollama contractually requires ZDR from their NCPs, while OpenRouter's free endpoints do not.

My assessment: Ollama Cloud is probably safer than OpenRouter's free tier, but it still sends your data to remote infrastructure. For truly sensitive work, local-only is the only guarantee.

My Setup

I run GLM-5.1 as my primary model. It has two modes:

• glm-5.1:cloud — runs on Ollama's cloud (NCPs with ZDR contracts)
• Local models — qwen3.5:9b, gemma3:12b, etc. run on my hardware

The cloud variant is fast and capable, but prompts travel to Ollama's servers. For day-to-day work that isn't sensitive, this is fine. For proprietary logic, we switch to local-only.

The Real Cost Comparison

Let me put actual numbers on this. Here's what you're trading when you use free vs paid models:

What does "fractions of a cent" actually look like?

Using paid ZDR models for private work on OpenRouter:

• A quick code review with Amazon Nova Micro: £0.0001 (one hundredth of a penny)
• A complex debugging session with Gemini 2.0 Flash: £0.001 (a tenth of a penny)
• A deep analysis with Claude 3.5 Haiku: £0.004 (half a penny)

You could run 200 private code reviews on Nova Micro before spending a single penny. The idea that you need to use free models to save money is a false economy — you're paying with your data instead of your wallet.

Actionable Advice

If you're using any free AI model endpoint, ask yourself:

1. What am I sending? Code? Agent conversations? Business logic? If it's anything you wouldn't publish on GitHub, don't send it to a free endpoint.

2. Who is the provider? Check OpenRouter's provider page. If there's no "No training" badge, they're training on your data. Period.

3. Is there a paid alternative? For most tasks, a ZDR paid model costs less than a penny. The cheapest ZDR models on OpenRouter (Nova Micro at £0.028/M tokens) are so close to free it barely matters — but your data is contractually protected.

4. Can you go local? If you have a GPU, even a modest one, running a 9B model locally handles 80%+ of daily tasks with zero data leaving your machine. Ollama makes this trivial.

The Bigger Picture

We're in a gold rush period for AI. Companies like NVIDIA are offering free inference for the same reason Google offers free search and Meta offers free social networking — the users are the product. Your prompts, your code, your questions, your agent workflows — all of it becomes training data for models that will eventually be sold back to you (or your competitors) as enterprise products.

There's nothing wrong with this exchange if you go in with your eyes open. Free models are great for:

• Learning and experimentation
• Public content creation (blog posts, tutorials, documentation)
• General questions that aren't proprietary
• Evaluating models before committing to paid tiers

But if you're running autonomous agents that have access to your codebase, your configs, your API keys — that's when free gets expensive. Your competitive advantage (how you build, what you build, the problems you're solving) becomes someone else's competitive advantage too.

Use free models for what they're for: public, non-sensitive work. For everything else, a fraction of a cent buys you contractual privacy protection. Or just go local — your GPU doesn't gossip.

The Risk Matrix — At a Glance

View full-size infographic

💡 Found this useful? 👉 Follow @Raf_VRS for more AI privacy insights that put you in control of your hardware. Stop Scrolling. Start Building. 👉 Support the work: ko-fi.com/rafvrs