Every model. One search box.
Filter across frontier and open-source models by modality, org, context, and price. Point at any model with a single string like anthropic/claude-opus-4.5.
| Modality | |||||||
|---|---|---|---|---|---|---|---|
| anthropic/claude-opus-4.5 | Anthropic | 94.2 | 82 | 320ms | 200k | $15.00 | textvision |
| openai/gpt-5.5 | OpenAI | 93.8 | 74 | 410ms | 400k | $12.50 | textvisionaudio |
| anthropic/claude-opus-4.1 | Anthropic | 93.1 | 78 | 340ms | 200k | $15.00 | textvision |
| openai/gpt-5 | OpenAI | 92.4 | 78 | 420ms | 400k | $10.00 | textvisionaudio |
| google/gemini-3.1-pro | 92.1 | 118 | 180ms | 2000k | $7.00 | textvisionaudiovideo | |
| openai/gpt-5-codex | OpenAI | 91.2 | 96 | 260ms | 400k | $4.50 | text |
| anthropic/claude-sonnet-4.5 | Anthropic | 90.8 | 118 | 220ms | 200k | $3.00 | textvision |
| xai/grok-5 | xAI | 90.6 | 95 | 260ms | 256k | $6.00 | textvision |
| google/gemini-2.5-pro | 90.4 | 108 | 210ms | 2000k | $5.00 | textvisionaudiovideo | |
| openai/o3 | OpenAI | 90.1 | 62 | 520ms | 200k | $10.00 | textvision |
| deepseek/deepseek-v4 | DeepSeek | 89.9 | 140 | 210ms | 128k | $0.55 | text |
| anthropic/claude-sonnet-4 | Anthropic | 89.4 | 122 | 210ms | 200k | $3.00 | textvision |
| alibaba/qwen3-coder-480bhosted | Alibaba | 89.4 | 130 | 220ms | 256k | $1.20 | text |
| xai/grok-4 | xAI | 89.1 | 90 | 280ms | 256k | $5.00 | textvision |
| alibaba/qwen3-235b-thinking | Alibaba | 88.9 | 82 | 340ms | 128k | $0.90 | text |
| alibaba/qwen3-max | Alibaba | 88.4 | 128 | 240ms | 1000k | $0.90 | textvision |
| moonshot/kimi-k2-thinking | Moonshot | 88.4 | 82 | 360ms | 256k | $3.00 | text |
| openai/gpt-4.1 | OpenAI | 88.2 | 92 | 300ms | 1000k | $8.00 | textvision |
| deepseek/deepseek-r1 | DeepSeek | 88.2 | 90 | 320ms | 128k | $0.55 | text |
| deepseek/deepseek-coder-v3 | DeepSeek | 88.2 | 155 | 200ms | 128k | $0.60 | text |
| anthropic/claude-sonnet-3.7 | Anthropic | 88.1 | 128 | 200ms | 200k | $3.00 | textvision |
| meta/llama-4-405b | Meta | 87.6 | 165 | 150ms | 128k | $2.40 | textvision |
| openai/o4-mini | OpenAI | 87.5 | 88 | 380ms | 200k | $1.10 | textvision |
| deepseek/deepseek-v3.2-exphosted | DeepSeek | 87.1 | 150 | 200ms | 128k | $0.28 | text |
| alibaba/qwen3-vl-235bhosted | Alibaba | 87.0 | 76 | 380ms | 128k | $2.00 | textvision |
How to read the catalog
Every row is a model as it is actually served, not as it was announced. Prices are USD per million tokens at provider list — llmcloud adds no margin, so the number you see is the number you are billed. Where several providers serve the same weights, the catalog shows the best currently available combination and the router picks between them per request.
The usable window under real concurrency, including KV cache headroom — not the architectural maximum quoted at launch.
Median decode tokens per second per stream, measured by our probes rather than taken from vendor marketing.
Time to first token at p50 from our edge. This dominates perceived speed in interactive products far more than total throughput.
A blended score across task-class evals, refreshed daily. Use it to shortlist, then confirm on your own traffic.
What the endpoint actually accepts today. A model with vision weights is only listed as vision if the served endpoint takes images.
Input and output rates listed separately because the blend depends entirely on your workload. Reasoning traces bill as output tokens.
Picking a model without guessing
Headline benchmark scores are the worst way to choose. Three numbers in this table predict production outcomes far better: blended price at your real input-to-output ratio, p50 time to first token from your users' region, and the model's failure mode on malformed tool calls. A model two points down on a leaderboard but half the price and 300 ms faster is usually the better production choice, which is exactly the tradeoff smart routing automates. For per-task evidence, read the live leaderboard, or the workload playbooks.
Model catalog FAQ
Does llmcloud mark up token prices?
No. Tokens are billed at provider list price with no platform fee and no credit float. Revenue comes from Team seats at a flat 10 USD per user per month and from hosted inference on our own hardware.
Why do two providers show different prices for the same model?
Because they run different serving configurations, hardware, and margins on identical open weights. That spread is the reason routing exists — the router picks per request based on live price, latency, and success rate rather than a fixed vendor preference.
How often is the catalog updated?
Prices and availability are polled from provider manifests continuously; latency and throughput figures refresh from probe data through the day; quality ranks recompute daily. New model launches are added once they pass the compatibility suite.
Are quantized builds labelled?
Yes. Served precision is a published field and providers agree during certification not to substitute a quantized or different build under a name implying the reference weights. Violations remove a provider from the pool.
Can I restrict my organisation to a subset of this catalog?
Yes. API keys carry allow-lists for models, routing policies, providers, and regions, so a staging key can be limited to cheap models while a production key is limited to zero-retention providers in a single jurisdiction.