The right AI for the job. The right prompt for it.
Act one · the map
Every model, placed by what it costs and what it can do
This is not an illustration. Each point is a real model from the catalog: left is cheap, right is expensive, higher is a stronger measured score, and the newest arrivals sit closest to you. Move the cursor to look around, or drag with a finger.
Act two · the ranking
A ranking is where most tools stop
The same points, sorted by the public index everyone quotes. It takes four seconds to read, it is genuinely useful, and it still has not told you what to do with the thing you actually have to write this morning.
Act three · your job
So start from the job, not from the leaderboard
Describe what you need to get done, in any language. The whole catalog collapses onto the models that fit it.
Act four · the rest of it
Everything else is built on the same two answers
Which model, and what to send it. Every tool below is one of those two questions asked in a different shape.
Describe your goal
WhichAI turns it into a prompt tailored to each AI model - and tells you which model is the best fit for the job, in plain language.
Describe your goal to generate prompts.
Uses your free API key to make the goal clearer and more complete - you review before applying.
Target models
AI assistants
Ecosystem assistants
Open models free/cheap via Groq & OpenRouter
Select at least one target model.
Your optimized prompts will appear here - one version per model, plus a recommendation of which model fits this task best.
1 Describe your goal
2 Copy the prompt tailored to each AI
3 Compare the answers and iterate
No account · No tracking · Nothing you write leaves your browser
Optimized prompts
Model guide
Which AI app fits each kind of task - public benchmarks translated into plain language.
Market at a glance
Top models - intelligence index
Price vs performance
Model catalog
Compare
Run the same optimized prompt in each AI app, paste (or auto-run) the answers, and score them side by side. Everything stays in your browser.
No active comparison. Generate prompts first, then click "Compare outputs" to bring them here. Go to the Generator
Saved comparisons
Merge studio
Read every answer side by side and hand-pick the best parts into one final draft. Nothing is automatic - you stay the editor.
No outputs yet - run or paste answers in Compare first.
Add another output
Chains
Break a complex goal into linked steps: each step gets its own optimized prompt, and the output of one step feeds the next. Auto-run steps where you have a free key; copy & paste elsewhere.
Describe your goal to build a chain.
Workflow roadmap
Saved chains
Settings
API keys enable automatic execution in Compare. They stay in this browser and are sent only to the provider you call, never to WhichAI's own API, which is read-only and accepts no data.
How to get free API keys
Gemini - open aistudio.google.com/apikey, sign in with Google, click "Create API key" and copy the key (starts with AIza…). No credit card.
Groq - runs Llama - open console.groq.com/keys, create a free account, click "Create API Key" (starts with gsk_…).
OpenRouter - open models - open openrouter.ai/keys, sign up free, click "Create key" (starts with sk-or-…). Model IDs ending in ":free" cost nothing.
Paste each key below, press "Test key" to verify it, then Save. Keys never leave this browser.
Where to keep API keys
"Session only" keeps the key in memory and forgets it when the tab closes. "Saved on this device" writes it to browser storage, where it survives a restart - and where any script running on this page could read it. That is why it stays opt-in.
Your keys are being written to this browser's storage. On a shared or work computer, prefer session only.
What a key can do, and what happens here
What the key can doAnyone holding it can spend your quota with that provider, under your account. On free tiers that means rate limits and a suspended key; on a paid account it means a bill. It cannot read your provider account, your other keys, or anything on this device.
Where it goes from hereStraight from this browser to generativelanguage.googleapis.com, api.groq.com or openrouter.ai - the only three destinations this page's Content-Security-Policy allows. WhichAI's own API is never in that path: it is read-only, takes no request body and would have nothing to leak.
Why not route it through a backendA proxy would mean your key sitting on someone else's machine and every prompt passing through it. Keeping the call in your browser is the more private design, not the lazier one - the trade-off is that browser storage is only as safe as the browser.
If a key leaksRevoke it and issue a new one - it takes a minute and costs nothing: Google AI Studio · Groq · OpenRouter. Never reuse a key you also use in a deployed app.
Good habitsUse free-tier keys only, one per tool, and set a spend limit where the provider offers one. WhichAI works fully without any key: copy and paste covers every model.
Google AI Studio - runs Gemini
Default: gemini-3.6-flash (free tier). Change only if Google renames its models.
Groq - runs Llama
Default: llama-3.3-70b-versatile (free tier). See console.groq.com/docs/models for alternatives.
OpenRouter - runs open models (Nemotron, Qwen, GLM, Kimi, DeepSeek)
One free key unlocks auto-run for several open models. Model IDs ending in ":free" cost nothing (about 20 requests/min, 200/day). Get a key at openrouter.ai - no credit card needed.
Defaults are free routes: nvidia/nemotron-3-ultra-550b-a55b:free and qwen/qwen3-coder:free. GLM, Kimi and DeepSeek currently have no free OpenRouter route - leave empty to keep copy & paste, or set a paid ID (billed to your own OpenRouter credits) from openrouter.ai/models.
Preferences
WhichAI automatically remembers the last options you used in the Generator (models, task type, format, length, tone) and restores them next time.
Your data
Everything WhichAI saves lives in this browser: preferences, comparisons, chains, merge drafts and (if you chose so) API keys. Export a backup before switching device or clearing the browser.
An open, free, sincere project with one goal: reduce the uncertainty of using AI - which model, which prompt, with evidence.
How WhichAI is different from a leaderboard
Leaderboards measure models. WhichAI decides what to do with them. It does not run its own benchmarks and does not pretend to: it stands on published evaluations, cites them with a date, and adds the layer none of them cover - going from a goal written in plain language to a model choice and a prompt that model responds well to.
Comparison between WhichAI and other AI model resources
LMArena · Artificial Analysis
Hugging Face
WhichAI
Question it answers
Which model scores higher?
Where do I get and run a model?
Which AI should do this job, and what do I send it?
What you leave with
A ranking
Weights, datasets, spaces
A model pick, the reason, and a prompt written for it
Made for
Researchers tracking the frontier
ML engineers and builders
Anyone with a task and twelve open tabs
Where numbers come from
Their own evaluations
Community uploads
Their published snapshots, cited and dated
Needs an account
No
For most things, yes
No, and nothing here stores one
If all you need is a ranking, use Artificial Analysis or LMArena directly - they are linked on every model page here. WhichAI is the layer on top, not a replacement.
The WhichAI API
The catalog is not locked inside the page. Every endpoint is read-only, needs no key, sends CORS headers and is cached at the edge, so anyone can build on the dataset under CC BY 4.0.
GET /api/healthStatus, data volume, snapshot dates and the list of endpoints.
GET /api/modelsThe full catalog with search, tag and vendor filters, sorting and pagination. Add ?id=claude-opus-5 for one model.
GET /api/benchmarksThe curated task rankings with sources. Add ?task=coding to run the router server-side.
GET /api/recommendSend ?goal= a plain-language goal: the server detects the task type and answers with a recommended AI plus a free alternative.
GET /api/statsAggregates computed on the server: totals, score distribution, tallies by vendor and tag, leaderboard.
Read-only by design: the API accepts GET and nothing else, takes no request body and stores nothing. Your goals, prompts and keys never reach it.
How the data stays current
Not by hand, and not by pretending a machine can judge a benchmark. The split is deliberate:
Automatic, every week - a scheduled job queries public sources, verifies that every free model route the app ships still exists, flags price and context-window drift beyond a threshold, and lists models released since the last snapshot. It then regenerates the static pages and the sitemap and runs the full test suite.
Automatic, on every change - continuous integration runs the static checks, a DOM smoke test that boots the real app, and the API tests. A change that breaks any of them never reaches the site.
Human, on purpose - intelligence scores are never rewritten by a script. A drifting number opens a pull request with the evidence; a person confirms it against a dated public snapshot and cites it. The whole site is judged on those numbers being honest, so a robot does not get to change them quietly.
Every page shows the snapshot date it was built from. When the API is reachable, the Model guide reports the live figures the backend is serving.
FAQ
Is WhichAI really free?
Yes, completely. No account, no subscription, no ads, no tracking. It runs on free hosting and the optional auto-run features use your own free API keys (Google AI Studio, Groq). There is nothing to sell you.
Where do my data and API keys go?
Nowhere. Everything (goals, prompts, comparisons, chains, keys) lives only in your browser. By default API keys are kept for the current session only; saving them on the device is an explicit opt-in in Settings. WhichAI runs a read-only API that serves the public model catalog and nothing else: it accepts no request body, stores nothing and never sees what you type. API keys are sent only to the provider you call (Google, Groq or OpenRouter), directly from your browser.
Why are generated prompts in English?
Every major model performs measurably best with English instructions. Each generated prompt includes a rule telling the AI to answer in the language of your task - so you can write your goal in any language and get the answer in your language, with English-optimized instructions doing the heavy lifting.
How reliable are the benchmark numbers?
Where public leaderboards publish a score (LMArena, SWE-bench, Artificial Analysis) we use it and link the source. Where they don't - private, preview, rumored or niche models - we show a clearly marked estimate ("~", "est."). Data is refreshed roughly monthly; the snapshot date is always shown.
What do the category scores and labels mean?
Each model in the database gets one overall score (Artificial Analysis Intelligence Index where published) plus four category ratings - Coding, Reasoning, Writing, Agents & tools - on a 0–100 scale. Category ratings are WhichAI blends of public category leaderboards and editorial judgment: use them to compare models at a glance, not as official measurements. Labels (like "free", "open weights", "in Generator") are filters: "in Generator" means you can generate optimized prompts for that model family right here.
Do I need API keys to use this?
No. Copy & paste works with every model. Keys only add convenience: with free Google AI Studio, Groq or OpenRouter keys, Compare and Chains can run Gemini, Llama and several open models (Nemotron, Qwen Coder…) automatically.
How does the model recommendation work?
A curated dataset translates public benchmarks into per-task rankings with a confidence level and linked sources. It is guidance, not gospel - near-ties are declared as ties, and the router never hides its reasoning.
Can I contribute?
Yes, in two ways. The dataset is open under CC BY 4.0 (/api/models, data/models.json) so you can build on it or check the numbers. And corrections are the most valuable thing you can send: a wrong score, a missing model, a bad translation, a screen that confused you. Contact details are below.
Similar tools - and what they do better
WhichAI is not the only prompting tool. In the spirit of openness, here is the honest landscape:
PromptPerfect - AI-powered engine that rewrites an existing prompt for maximum quality, including image prompts (Midjourney, Stable Diffusion).Better at: one-click automated optimization of a prompt you already have (paid credits).
AIPRM - Chrome extension with 5,400+ curated prompt templates inside ChatGPT.Better at: ready-made templates with one click in the browser, if you live in ChatGPT (from $20/mo).
PromptBase - marketplace of expert-crafted, human-reviewed prompts ($1.99–9.99 each).Better at: buying one proven prompt for one specific job, with example outputs.
PromptLayer - Git-style prompt versioning and observability for teams.Better at: professional prompt lifecycle management for developer teams shipping LLM apps.
Anthropic Console & OpenAI Playground - the vendors' own prompt-improvement and testing tools.Better at: testing with real API calls and vendor-official optimization for that one model.
What WhichAI offers instead: prompts for 13 models from a single goal, benchmark-based model recommendations, side-by-side comparison and multi-step chains - 100% free, no account, everything in your browser.
Methodology v1.0
Updated: July 20, 2026 · next review with the monthly data refresh.
Overall score - Artificial Analysis Intelligence Index (2026 rebased scale), taken from one dated public snapshot per refresh so every model sits on the same scale. Models without a published score get a clearly marked estimate ("~", "est.") and are never ranked against measured scores.
Category ratings (0-100) - WhichAI blends of public category leaderboards (LMArena coding, SWE-bench, agentic and writing boards) plus editorial judgment. Guidance, not measurements.
Task router - per-task rankings built from the benchmarks above, with a confidence level (high / medium / low) and linked sources. Near-ties are declared as ties.
Prices, context, speed - vendor pages and public pricing mirrors, with the retrieval date stored in the database. Unpublished values show "n/a" and never enter charts.
Status labels - public, preview, private, legacy, rumored. Rumored and private models are info-only: they never appear in rankings or recommendations.
Independence - no sponsored placements, no affiliate rankings, no accounts, no tracking. Corrections are welcome and get fixed in the next data refresh: the contact details are in Contribute below.
Limits: benchmarks are partial proxies of real work; snapshots age between refreshes; category blends involve judgment. When in doubt, run your own comparison in Compare.
Support WhichAI
WhichAI is free, has no ads and no sponsors: the only costs are the domain and the hours that go into keeping 100+ models honest and up to date. If it saved you time, a small donation keeps it independent.
Donations never influence rankings: methodology and sources stay public.
Contribute & contact
The dataset is open (CC BY 4.0) and reachable through the API; the application source is not public. Wrong data, missing models, bad translations and confusing screens are the most useful things you can send:
Open dataset - /api/models and data/models.json109+ models with scores, prices and sources under CC BY 4.0. Attribution: WhichAI, whichai.wiki.