AI model comparisons

Each lab’s most capable models, for the hardest work.

Claude Fable 5.1 vs GPT-6 Astra1,000 requests: $80.00 and $80.00Claude Opus 5.5 vs GPT-6.1 Sol1,000 requests: $32.00 and $16.00Gemini 3.1 Pro vs Claude Opus 5.51,000 requests: $18.00 and $32.00GPT-6.1 Sol vs GPT-6 Astra1,000 requests: $16.00 and $80.00

The models most teams run day to day, where price and quality are both in play.

Claude Sonnet 5.5 vs GPT-6.1 Sol1,000 requests: $16.00 and $16.00Claude Sonnet 5.5 vs Claude Opus 5.51,000 requests: $16.00 and $32.00Claude Opus 5.5 vs Claude Opus 51,000 requests: $32.00 and $40.00Claude Sonnet 5.5 vs Gemini 3.8 Flash1,000 requests: $16.00 and $6.00Grok 4.6 vs GPT-6.1 Sol1,000 requests: $24.00 and $16.00GPT-5.3-Codex vs Claude Sonnet 5.51,000 requests: $19.25 and $16.00

Small models for high volume: classification, extraction and quick answers.

GPT-6 Luna vs Claude Haiku 4.51,000 requests: $0.80 and $8.00GPT-6 Luna vs Gemini 3.8 Flash1,000 requests: $0.80 and $6.00DeepSeek-V4.1-Flash vs GPT-6 Luna1,000 requests: $1.36 and $0.80GLM-5.3-Flash vs Qwen 3.8 Flash1,000 requests: $1.24 and $1.02

Models whose weights are published, compared with each other and with closed rivals.

DeepSeek-V4-Pro vs Kimi K31,000 requests: $7.80 and $29.64DeepSeek-V4-Pro vs GLM-5.31,000 requests: $7.80 and $7.83Qwen 3.8 Max vs DeepSeek-V4-Pro1,000 requests: $13.26 and $7.80XMiMo-V2.5-Pro vs Kimi K31,000 requests: $2.83 and $29.64Kimi K3 vs Claude Sonnet 5.51,000 requests: $29.64 and $16.00gpt-oss-120b vs Llama 3.3 70B1,000 requests: $1.36 and $4.07Qwen 3.8 27B vs Gemma 4 31B1,000 requests: $4.09 and $1.07Qwen 3.8 27B LK vs Qwen 3.8 27B1,000 requests: $2.40 and $4.09

Models that turn text into vectors for search and retrieval.

text-embedding-3-large vs text-embedding-3-smallInput: $0.13 and $0.02 per 1M tokens

Each cost is the same job on both models: 1,000 requests of about 3,000 tokens in and 1,000 out, at Roar AI’s published rates. Models that think before answering spend extra output tokens doing it, so on hard prompts the gap can be wider. Embedding models are compared on their input rate. For every model’s full rates, see the model directory.