AI Lifter
Get the app
FIG. 10 · The lineup
Gym Bro, the optional AI Coach, runs on your own DeepSeek, Gemini or OpenRouter key — so which model powers it is up to you. We've driven the whole shortlist through the app's real coaching scenarios — 58 questions, the same persona, tools and rules the app ships — to see what each one is actually like in the coach's chair. Six we'd use; one we'd steer around.
These are our dev-time findings, nothing more. The in-app picker stays brand-neutral and never ranks by name.
01 The shortlist
Every one answered the same 58 real questions in the coach's seat. Prices are input / output per million tokens, from the in-app picker. All of them are good enough on the content — the differences below are about delivery, caution and cost.
~deepseek/deepseek-pro-latest
Clean, accurate and even-handed — the baseline every other model is measured against. A little less eager to reach for tools than the rest, which mostly reads as calm. Text-only: it can't read a food photo.
google/gemini-flash-latest
Fifty-eight for fifty-eight clean, fast, and it reads images. None of the reasoning-leak its bigger sibling has (below) — the light-reasoning tier simply doesn't spill. A strong, inexpensive everyday pick.
~openai/gpt-mini-latest
Clean and the most concise of the lot — short, direct answers that fit a phone. It enforces OpenAI's strict tool contract, so it wouldn't talk to the app at all until we tightened every tool schema (see below).
~openai/gpt-sol-latest
The most safety-thorough of the five — it volunteers the red flags the others don't (dark-urine rhabdo on bad soreness, cervical-spine caution for rheumatoid arthritis). More thorough also means more verbose, and it's the priciest seat in the house.
google/gemini-3.5-flash-lite
Clean across the board and good at reading your log — none of its bigger sibling's reasoning leak. Tool-eager like the other Geminis (it'll offer to remember a health fact), and a hair less careful than GPT Sol on one medical hand-off, but accurate and safe throughout.
~deepseek/deepseek-flash-latest
The cheapest seat here by a mile, and the sharpest at using your data — it builds answers around your actual set counts and session times more than any other model. The catch: delivery wobbles now and then — a stray line of thinking, and once a fully garbled reply — so it's less polished than Gemini Flash or V4 Pro. Strong value if you can live with the occasional hiccup.
google/gemini-pro-latest
The content is as good as the rest — but about one reply in four leaks its raw chain-of-thought or tool-call JSON into what the user sees, and one in ten comes back unusable. It's a quirk of this model, not the app (the adapter routes reasoning to a hidden channel; Pro writes it into the answer anyway). Right now gemini-pro-latest resolves to the snapshot codenamed Argon, so the leak is Argon's — it could shift if the alias is repointed. Same price as GPT Sol, for a worse experience.
02 Things worth knowing
Every model runs on a key you bring. AI Lifter never resells access or marks it up, and the in-app picker lists models neutrally — these notes only inform what the team would reach for.
Across 58 questions, DeepSeek V4 Pro, Gemini Flash, GPT Mini, GPT Sol and Gemini 3.5 Flash Lite landed on the same answers (DeepSeek Flash agrees on content too — it's only the delivery that wobbles). The coach's character comes from its persona and skills, not from which model you point it at.
Gemini Pro spills chain-of-thought and tool JSON into roughly a quarter of its replies — stray tokens, half-finished reasoning, the occasional non-answer. Upstream, not us; the fix is simply to not use it.
GPT Mini and GPT Sol enforce the strict function-calling contract to the letter. They rejected the app's entire tool set until we rewrote every tool schema to satisfy it — a real fix, shipped.
Photo food-logging needs a model that can see. DeepSeek V4 Pro is text-only, so the app tells you up front when your chosen pair can't read an image rather than failing mid-task.
From a couple of cents per million tokens at the cheap end to $2 in / $10 out for the premium seats — for the same coaching job. Cheaper is often the smart default; the coach barely notices.