The Best Model For
Which AI model to use, from ten labs. Pick a job, see what it costs today, and when a cheaper model will do.
Short answer
For a normal coding job, start with Claude Sonnet 5.5 ($2.00 in · $10.00 out per million tokens) and read why a flagship is usually wasted. The catalogue has 290 current rows across 10 labs, 204 of them chat models.
Start from the job
The ten labs
Cheapest and longest are among chat models only, so an embedding row or an unpriced video model does not win the column. Open weights means the feed lists a public weights id.
| Lab | Rows | Open weights | Cheapest chat input | Longest chat context |
|---|---|---|---|---|
| OpenAI | 79 | 5 | gpt-oss-20b $0.018 in · $0.09 out | 1,050,000 GPT-5.4 |
| Anthropic | 16 | 0 | Claude Haiku 5.5 $0.10 in · $0.50 out | 1,000,000 Claude Fable 5 |
| Google (Gemini) | 49 | 6 | Gemma 3 12B $0.05 in · $0.15 out | 1,048,576 Gemini 2.5 Flash |
| xAI (Grok) | 12 | 0 | Grok 4.20 $1.25 in · $2.50 out | 2,000,000 Grok 4.20 |
| Meta (Llama) | 18 | 9 | Llama 3.1 8B Instruct $0.02 in · $0.04 out | 1,310,720 Llama 4 Scout |
| Mistral | 30 | 17 | Mistral Nemo $0.029 in · $0.03 out | 1,048,576 Mistral Large 4 |
| DeepSeek | 14 | 14 | DeepSeek V4 Flash 0731 $0.0443 in · $0.30 out | 1,048,576 DeepSeek V4 Flash 0423 |
| Alibaba (Qwen) | 52 | 32 | Qwen3.7 Flash $0.03 in · $0.13 out | 1,048,576 Qwen3.8 2.4T A95B |
| Cohere | 12 | 5 | Command R7B (12-2024) $0.0375 in · $0.15 out | 256,000 Command A |
| Amazon (Nova) | 8 | 0 | Nova Micro 1.0 $0.035 in · $0.14 out | 1,000,000 Nova 2 Lite |
What changed recently
How to read a price
$2.00 in · $10.00 out means two dollars per million input tokens and ten per million output tokens. A short email is a rounding error. A long contract with a long answer is not.
A published 0 means that field is free, per OpenRouter's docs. It does not mean every other meter is free. Unknown means the field was missing. We do not guess.
Questions
Where do the prices and context windows come from?
From OpenRouter's public models API, which needs no key. Each fact on a model page names the field it came from and the day The Best Model For last checked. A missing value is shown as unknown. Nothing is filled in.
Do you run benchmarks?
No. If the feed includes a third-party index from Artificial Analysis, it is shown with that attribution. The Best Model For does not invent scores, and the guide notes are opinions about fit, not measurements.
How do I pick a model if I am not technical?
Start from the job, not the brand. Open the task finder, choose the work and how much the bill matters, and read the short answer. Then open the model page and check the price against a cheaper one.
How fresh is this?
A scheduled job refetches the public feed every day, records models that were added, removed, or repriced, and the pages are rendered from that snapshot. The date in the header is the check, not a guess.
Is The Best Model For sponsored by any of these labs?
No. The guide is independent. Links to labs and to OpenRouter are citations, not endorsements.