The Expensive AI Apologized Its Way Out of a Raise — Then Its Newest Version Fixed It

Can a cheap AI coach a real salary negotiation? Thirty job offers, three models, one blind judge — run on the newest version of each model. The cheapest one won:

30 offers per model, judged blind by a separate AI. Model versions as of Aug 24, 2026. “0.95+” = nothing meaningful to deduct.

What we did

We built a Salary Negotiation Coach: give it the offer, your experience, your leverage, and your biggest worry. It writes the negotiation email, a 60-second phone script, replies to the three most likely objections, and a realistic ask.

The 30 test scenarios ranged wide: a nurse, a teacher, a retail shift manager, a staff engineer. Hourly wages and $200k+ packages. Half had no competing offer. Worries included “I hate conflict”, “I’m on a visa”, and “I was told the band is fixed”.

A separate model judged each plan blind on four things: grounded (no invented market data — automatic −0.5), assertive-but-not-adversarial, specific to this person’s leverage, and a defensible ask with a fallback.

A real output

What we learned

1. Versions matter as much as vendors.
Swapping gemini-3.5-flash for 3.7-flash — nothing else changed — moved its score from 0.68 to 0.75, cured the apologetic tone, and cut its price 3.5×. If you benchmarked a vendor six months ago, that result is stale.

2. Tone failures are directional.
Old flash failed soft — polite surrender. Haiku’s failures lean the other way, with asks that go weak or a rationale that overreaches. Luna held the middle: warm and firm.

3. The winner never made things up.
The rubric’s biggest penalty was for inventing market data. Luna took zero such penalties across 30 plans; its worst miss was hedging instead of naming a firm target.

4. Cheap won on every axis, twice.
Luna was 8× cheaper than the new flash, fastest, and highest-scoring — the same sweep it pulled on our dating-bio benchmark.

The fine print

• Judge: gpt-5.4-mini — not a contestant — scoring 0–1 blind against the rubric.
• Sample: 30 scenarios × 3 models per round, zero failures. Run-to-run judge variance is real (±0.05–0.1 on the same model), so trust the ranking, not the second decimal; the table shows one round on the latest versions.
• Cost: both rounds together, judging included, about $3.40.

Run this on your own use case

The flow was described in one sentence in Evaligo; each experiment variant swaps the model and nothing else — which is also why re-running on a new model version takes minutes. If your product depends on tone, this is the benchmark to run before you pick a model, and again every time a vendor ships an upgrade.