When you click a "Try it on" link from a model page, you end up on one of these. Here's why we link to them, what we like about each, and how we make money.
One API key, hundreds of models, per-provider pricing.
OpenRouter is the backbone of our pricing column. Their/api/v1/modelsendpoint exposes prompt and completion costs per million tokens, image-generation prices per unit, and the full list of underlying providers serving each model. We sync that every day and join it against LMArena's rankings.
What we like:
The advertised price is what the upstream provider charges - OpenRouter takes a small credit-card fee, not a per-token cut.
Every model lists its providers with uptime, latency, throughput, quantization, and context length. If one is down or throttled, traffic shifts automatically.
Switching from Claude to GPT to Llama is a model-id change, no new account or SDK to wire up.
Free tiers, BYOK variants, and provider-routing preferences are all surfaced cleanly in the API.
You pre-pay; when credits run out, requests fail. No AWS or Google Cloud-style surprise bills from a runaway script. Auto-top-up exists but you have to opt in.
The image/video models OpenRouter doesn't carry.
Most of the interesting image and video generation models on LMArena's leaderboards aren't on OpenRouter, they live on specialized hosts. Wavespeed is the one we've found that covers the widest slice of them (Seedream, Veo, Wan, Kling, grok-imagine, gpt-image-2 variants, and more) with consistent per-run pricing.
What we like:
Without Wavespeed, the video subset of our leaderboard would be a long list of models you can't actually run anywhere. They turn arena-only entries into clickable links.
Their variants expose a base price and (when it's conditional on resolution, duration, or quality) a formula we can decipher. That's how we display a real default-call cost on the leaderboard.
Same key pattern across every model they host: generate, poll, fetch result.
When a new model debuts on LMArena, Wavespeed usually has it within a few days, which is what lets our "try it on" links keep working as the field moves.
Same as OpenRouter: pre-paid credits, no card-on-file surprises. Auto-top-up is opt-in, off until you flip it. Good for testing a model without worrying about a runaway script.
Between OpenRouter and Wavespeed we can link to a host for almost every model on the leaderboard. We don't add a provider just because it exists, only when it covers models the others don't, or solves a real problem (per-run pricing, arena coverage, etc).
Wavespeed links carry a ?ref=gptbased referral parameter. When someone clicks through from gptbased and signs up, we get a referral credit on Wavespeed. That's the only revenue this site currently earns.
OpenRouter links are plain. They don't run an affiliate program, and we don't get anything if you click through.
We chose to add Wavespeed because they cover image and video models that nothing else does. The referral didn't influence that decision, and it doesn't change rankings, sort order, or anything else on the leaderboard. The order of "Try it on" links is fixed (OpenRouter first when both are present), regardless of payout. Never paid placement.
If a provider deserves a spot, we'll add it. Drop us a line via the contact page.