Compare AI models side by side
Model Compare runs your own prompt on up to six models at once, from Claude and GPT to Gemini. Watch the answers arrive with how soon each one starts, its speed and its cost, vote blind, and let an AI judge rank them.
New to this? What it does, in plain words
It sends the same request to several AI models at once and shows their answers next to each other.
Use it when you need to choose a model for a job, or want to know whether a cheaper one is good enough. A finance team can see which model summarizes a quarterly report most accurately for the least money. A support lead can compare how each model answers the same tricky customer question, and vote before seeing which model wrote which answer.
What you need
Running a comparison needs an API key for each provider whose models you pick. Google AI Studio, OpenRouter and Groq give free keys. Without a key, open the example to see what a finished comparison looks like. Your keys, prompts and answers stay in this browser and go only to the providers you pick.
How to use it
- Type your prompt. Open System prompt if every model should get the same standing instructions first.
- Under Models, choose two to six models.
- Press Run. Each answer fills its own column, with its time, token count and cost underneath.
- Turn on Blind vote to pick the best answer before the names appear, or run the AI judge to score the answers against criteria you write.
Results
ExampleSummary
Example Select a heading to sortAI judge
OptionalA model reads the answers as A, B, C… in a random order, without names, and scores each one against your criteria.
Your arena record
Comparing models fairly
The same prompt goes to every model at the same moment, straight from your browser to each provider, with your own keys.
- Free, in your browser. Write one promptThe same instructions, variables and output limit for every model, so the race is fair.
- With your key. Up to six models answer at onceSecondsAnswersEach reply streams into its own column.SpeedTime to the first word, and words per second after it.CostWhat each answer cost, from published prices.
- Free, in your browser. Vote blindHide the names and pick the better answer. Your votes build a record of what you prefer.
- Free, in your browser. Keep the winnerSave the prompt to My Prompts, or test it on real cases in the Prompt Tester.
Make the comparison fair
- Same prompt, same settings. Every model gets the identical system prompt and prompt, with your variables filled in the same way, and the same output limit.
- Run it more than once. Answers vary from run to run. Rerun a model, or the whole comparison, a few times before you decide from a single answer. Every run is kept in History.
- Know what temperature and effort do. Temperature changes how varied the wording is. Many current models ignore or reject it, including the Claude 5-series, OpenAI's reasoning models and Gemini 4, so the arena leaves it out for them. Effort sets how long reasoning models think: more effort is usually better on hard tasks, and slower and costlier. Compare at the effort you would use in production.
- Vote blind. Names change what we see. Blind vote hides them and shuffles the columns until you pick the best answer, a tie, or "all bad". Then the names are revealed.
- Treat the judge as a second opinion. The AI judge reads the answers labeled A, B, C… in random order and scores them against your criteria. Judges tend to favor longer answers, the first answer they read and models from their own family, so pick a judge from a different family, make the criteria specific, and read its reasons. A high score can hide an invented fact.
How speed and cost are measured
- Starts in is the time to first token: from sending the request to the first visible text. It is measured in your browser, so it includes your network.
- Total runs until the last token arrives.
- Speed is output tokens per second from the first token to the last. Reasoning models count their hidden thinking as output tokens but produce it before the first visible token, so their speed can read high. Total time is the fairer comparison for them.
- Tokens are the counts the provider reports. A ~ marks an estimate, used when a provider does not report counts.
- Cost is tokens times the price per million tokens in our catalog, checked on October 20, 2026. OpenRouter prices load live from its public model list. Cached input is priced at the cache rate when the provider reports it. Your provider's bill is the final word.
- Best value is the cheapest answer whose judge score is within one point of the top score.
Privacy
- Your prompt goes from this browser straight to each provider you pick, with your key. thinkhard.si never sees prompts, answers or keys.
- Comparisons and your arena record are stored in this browser only. Delete them from History, or everything at once in Settings.
- Share links carry the prompt, and the answers if you choose, in the part of the link after #, which browsers do not send to any server.
Questions
Is it free?
Yes. You pay each provider for the tokens you use with your own key, and free keys from Google AI Studio, OpenRouter and Groq work here.
Is "Your arena record" a leaderboard?
No. It is built only from your own blind votes on your own prompts, in this browser. That makes it a record of which models you preferred for your work, which can be more useful than a public ranking, but a handful of votes is mostly noise.
Why did one model fail when the others worked?
Each column shows the provider's reason: a missing or rejected key, a rate limit, an exhausted quota, or a model the key cannot use. Retry runs just that model again with the same prompt.
Why is an answer cut off?
Answer limit in Settings caps every answer, in tokens. Raise it and rerun the model. Reasoning models get extra room for thinking on top of the cap.
Can I compare one model at two settings?
Not in one run: each column is a different model. Run the comparison once per setting and compare the runs in History.
Does the judge see which model wrote what?
No. It sees the prompt, your criteria and the answers labeled by letter, in a shuffled order. It is told to treat the answers as material to evaluate and to ignore any instructions inside them.