Small vs Frontier
Ask one question and three models answer side by side, from a 1-billion-parameter model small enough for a phone to a frontier model, with the time and cost of every answer.
Llama 3.2 1B
1 billion parameters
small enough to run on a phone
Qwen3-30B-A3B
30B total · ~3B active
open model, mixture of experts
Claude Sonnet 5
frontier scale
size not disclosed
EXPECTED
What you’re looking at
“Which AI should I use?” usually gets answered with benchmark tables. This is the hands-on version: one question goes to three models at the same moment, and you watch the answers arrive.
- Llama 3.2 1B (Meta): about a billion parameters, small enough to run on a phone.
- Qwen3-30B-A3B (Alibaba): 30 billion parameters, but only about 3 billion switch on for each word, so it runs about as fast and cheaply as a small model.
- Claude Sonnet 5 (Anthropic): a frontier model. Its size isn’t published, but it’s in a different league.
Each column shows how long the answer took, how many tokens it wrote, and what it cost. The presets come with the expected answer, so you can grade them yourself.
What happened when I ran the presets
- “How many r’s in strawberry?” Only Claude said 3. Both smaller models said 2. That isn’t carelessness: models read tokens, chunks of letters, not individual letters (the Tokenizer Lens shows this), so counting letters is surprisingly hard.
- The bat-and-ball question. The 1B model gave the intuitive wrong answer, $0.10. Qwen and Claude both got $0.05.
- は vs が. This is where size showed most. The 1B model wrote a fluent, confident explanation that was simply wrong: it called を “a polite particle.” Qwen gave a thin but correct answer. Claude gave the explanation I’d want a student to read.
- Facts and haiku. All three named Marie Curie, Physics, 1903, and all three wrote a respectable haiku. For fluent writing and famous facts, small models are often good enough.
Your runs will vary: these models don’t give the same answer every time.
The real lesson: cost and speed
The small models often answer first, sometimes in under a second, and always for a tiny fraction of a cent. Claude takes one to four seconds and costs roughly a hundred times more per answer. Still well under a cent, but it adds up across millions of requests.
So the practical question isn’t “which model is best?” It’s which is the smallest model that’s reliably right for this job? That’s exactly how the rest of this lab is built. The SAT Question Studio uses a cheap model only to explain answers that code has already verified. Explain This Sentence uses Claude’s smallest model. The heavyweight is saved for essay feedback and the agent.
Guardrails
Each visitor gets 10 comparisons a day, requests pass a Cloudflare human check, and the lab has a daily spending cap. Questions aren’t stored by this site.