Bring Your Own AI Key: Same Fact-Check Rules on Gemini, DeepSeek, and Qwen

Every fact-check costs someone money in model calls. Grounnel now lets you bring your own API key for Gemini, DeepSeek, or Qwen, so your provider account pays for the model work. Web search stays on Grounnel's account.

The short version

• Qwen is the cheapest by far. A typical 10-claim check costs about $0.005, roughly a third of Gemini's cost and a fifth of DeepSeek's off-peak cost (a tenth at peak).
• Gemini scored highest. On our 28-case trap set, run three times each, Gemini matched 98 of 99 expected verdicts, DeepSeek 93, and Qwen 89.
• None of the three passed a known-false claim as true. Qwen made the only false accusations: two, both on one invented claim.
• The same rules apply to every model. A verdict must cite a real sentence from a fetched page, or it gets downgraded.

If you want the lowest cost, use Qwen. If you want the best results on this test set, Gemini costs about three times more. DeepSeek costs more than Gemini and scored lower, so on this test it isn't worth the extra. Use it only if you already have a DeepSeek account, and run it off-peak.

Your text is not private. Every check is saved to a permanent, shareable page that anyone with the link can open, including your text, the results, and the token counts. Your text is also sent to the provider you choose, under that provider's own data terms. Gemini's free tier, for example, may use it to improve Google's products. Don't use Grounnel for confidential material.

What your key pays for

A check has three stages:

1. Read: the model pulls the factual claims out of your text.
2. Search: Grounnel finds and fetches web pages about each claim.
3. Judge: the model gives each claim a verdict and points to the numbered sentences it relied on. Grounnel looks those numbers up and shows the actual sentences.

Reading and judging run on your key. Search stays with us, because the judging rules need pages split into numbered sentences so every citation can be checked against real text. A check covers up to 40 claims, about 3,000 characters. Own-key checks get 30 checks an hour instead of 5.

What happens to your key

• It's sent over HTTPS in a request header, never in the URL or body.
• In your browser, it lives only in the page's memory, never in storage.
• On our server, it's used for that one check and never stored, hashed or otherwise.
• Anything that looks like a key is scrubbed from error messages before they're logged.
• A quick test confirms the key works first. For Gemini and Qwen, that's a tiny model call worth a handful of tokens.

Use a separate key for Grounnel so you can revoke it at any time. The About page has the details.

What a check costs

After an own-key check, the result page shows one line like this, from a real 10-claim check on Qwen:

This check used ~96,021 input / 5,255 output tokens on your key.

A check makes many model calls, which is why the input count is large. That line measures usage. The dollar figures below are our estimates: measured tokens times each provider's list prices.

DeepSeek's balance actually dropped by only $0.37, not $0.58, most likely because repeated prompts were cached.

Method: measured October 29, 2026, at list prices per million input/output tokens: qwen3.7-flash, Singapore, $0.03 / $0.13 (every request stayed under its 32K-token tier); gemini-2.5-flash-lite, $0.10 / $0.40; and deepseek-flash, $0.15 / $0.60 off-peak and $0.30 / $1.20 peak, uncached. The DeepSeek test runs were done off-peak, so its peak figures are the off-peak ones scaled by the price difference. The 10-claim figures are scaled from the average per claim, and only the Qwen 10-claim example above was measured directly. Prices move, so treat this as a snapshot.

Same rules, different models

Our test set has 28 short cases, mostly traps: a claim that's almost true, a misleading headline, a "first" that was actually the fourth. We ran each case three times per model. The 33 claims with a known expected verdict give 99 scored checks per model. That's fewer than the cost table's claim counts, which include every claim a model extracted.

The misses fall into four kinds:

• False accusations: only Qwen. It twice called an invented claim, about a "Zorgonian Institute" expedition, false. The correct verdict is "can't verify".
• "Can't verify" on a checkable claim: DeepSeek did this three times on "The Great Wall of China was not built by the Roman Empire", and Qwen once.
• Errors: Gemini's one miss was a check that failed with an error.
• Claims not extracted: sometimes a model didn't pull a claim out of the text at all.

What doesn't differ is what happens after a model answers. Every verdict passes through the same code checks, which we call gates. The simplest is an evidence requirement: a "supported" or "contradicted" verdict must cite a valid sentence from a page Grounnel fetched, or it's downgraded to "unsupported".

When the gate rejected a correct answer

When we first connected DeepSeek and Qwen, both correctly judged "Buzz Aldrin was the first man to walk on the Moon" as false. But they didn't cite a sentence, because our integration wasn't asking for citations in a form they reliably followed. The gate downgraded every one of those verdicts anyway.

We didn't loosen the gate. We fixed the integration. In this benchmark, that claim came back "contradicted" with a valid citation in all three runs on both models. Grounnel will sometimes miss a correct verdict rather than pass one that hasn't met the evidence bar. That trade-off is deliberate.

What these results don't prove

• 28 cases is a small set, and rankings between models are noisy. This isn't a leaderboard.
• An evidence requirement isn't understanding. A clickbait headline, "Microsoft is actually responsible for creation of Apple's iPhone", was a figure of speech. In 74 earlier runs of that case, a model twice took it literally and called the true claim "Microsoft did not create the iPhone" false. The gate can't catch it, because the cited sentence is real.
• Costs are list-price estimates, not your actual bill.

Try it

Paste some text at grounnel.n30sk.cc, turn on "Run this check on my own AI account", and pick a provider. I'd especially like to hear where a verdict is wrong, where the evidence is weak, or where Grounnel missed something.