Which AI model should I use?

JobHunt asks you to bring your own model, which is a real question with a real cost attached. We tested five so you don't have to.

Setting it up

1. Create an account at openrouter.ai and add a few dollars of credit.
2. Create an API key and copy it.
3. In JobHunt: Settings → AI, set Provider to OpenRouter, paste the key, click Fetch Models, and choose google/gemini-3.7-flash.
4. JobHunt will ask you to confirm before it sends anything to a cloud provider. That prompt is deliberate — see below.

You can change this later at any time, and changing it doesn't rescore the jobs you already have unless you ask it to. It's a low-stakes decision.

Direct, or through OpenRouter?

You don't need OpenRouter. JobHunt talks to Google, OpenAI, Anthropic and DeepSeek directly — pick the provider in Settings → AI and paste that company's own API key. For the recommended model that means a key from aistudio.google.com.

Go direct if you'd rather one fewer company saw your résumé. OpenRouter sits between you and the model, so your employment history passes through it on the way. Going straight to Google removes that hop, and you're billed by Google rather than pre-buying credits.

Use OpenRouter if you want to try models. One key reaches all of them, so changing your mind is a dropdown rather than a signup — and JobHunt's live spend estimate reads OpenRouter's pricing feed, so it can only show you a running cost there.

On price the two are a wash, at least at list: OpenRouter passes provider rates through without a per-token markup, so Gemini 3.7 Flash is $0.75/$3.75 per million tokens either way. OpenRouter's fee is 5.5% on credit top-ups; against that, going direct is fractionally cheaper. Discounts muddy this in both directions and never last — OpenRouter is running 50% off this model until 27 August 2026, and Google's own introductory rate doubles on 1 January 2027. Pick on privacy and convenience, which are stable, rather than on a price gap that isn't.

What we tested

Two things. Judgement: does it credit experience your résumé doesn't actually support, or penalise a requirement nobody could fail? Consistency: does the same request twice give the same answer? Every model below saw the same tests in the same run, five times each.

Model	Judgement	Same answer twice?	100 jobs	500 jobs
Gemini 3.7 Flash	25 of 25	always	$1.42	$7.10
GPT-5.6 Sol	25 of 25	always	$7.31	$36.55
DeepSeek V4 Flash	30 of 33	always	$0.17	$0.85
GPT-5.6 Luna	23 of 25	1 verdict moved	$0.29	$1.47
Gemini 3.1 Flash Lite	18 of 25	1 verdict moved	$0.41	$2.05
Local model	not tested	—	free	free

Costs are estimates at each model's standard published rate, not invoices. Longer postings and retries cost more. Two things move these numbers a lot: aggregators run temporary discounts (Gemini 3.7 Flash is half price on OpenRouter until 27 August 2026, which is not what the table shows), and Gemini's own introductory pricing doubles on 1 January 2027. DeepSeek is served by ~30 different hosts at rates spanning 7×, so yours may differ. JobHunt shows a live estimate for your actual corpus in Settings → AI.

Two to avoid

Gemini 3.1 Flash Lite. Given a candidate who has run GPU workloads but never written CUDA, it credited "expertise in the CUDA ecosystem" as fully met on five runs out of five. That is exactly the flattery that makes a fit score useless. Its scores also swing wildly on identical input.

DeepSeek V4 Flash. It judges well and it's the cheapest thing here, but it was about five times slower than everything else and once returned a response JobHunt couldn't read at all. Workable if cost dominates; you'll notice the wait.

Why consistency matters

Hosted models are not deterministic, even at temperature 0 — the same request can come back with a different answer, because your request is batched with other people's and may be routed to a different machine. We've measured a job's score moving by up to 20 points between identical requests on weaker models.

That's enough to reorder your list around a threshold, which is why we weight it as heavily as accuracy. The two models at the top of the table didn't move at all.

If you'd rather pay nothing

Run a model on your own Mac with LM Studio or Ollama, and point JobHunt at http://127.0.0.1:1234. It costs nothing, and no job description or résumé ever leaves your machine.

The honest trade-off: local models are slower, and on consumer hardware they are generally less reliable at the structured extraction JobHunt asks for — a malformed response means a job fails to extract and gets retried. We have not benchmarked local models, so we can't tell you how much accuracy you give up. If privacy is the priority, this is the option; if accuracy is, it isn't.

What gets sent

With a cloud provider, JobHunt sends the job description text for extraction, and the description plus your résumé text for fit scoring. That's your employment history going to a third party, which is why JobHunt asks for explicit consent the first time and won't send anything before you agree. With a local model, none of this leaves your Mac. See the Privacy Policy.