LLM costs
| Provider | Model | Input Price | Output Price | Context Length | Output Length | Model Size |
|---|---|---|---|---|---|---|
| Alibaba | Qwen2.5-Max | $1.6 | $6.4 | 32k | 8k | |
| Azure | phi-3.5 | $0.00013 | $0.00013 | 1k | 1k | |
| Anthropic | Claude 3.5 Haiku | $0.8 | $4 | 200k | 8k | |
| Anthropic | Claude 3.5 Sonnet | $3 | $15 | 200k | 8k | ≈175B |
| Anthropic | Claude 3 Haiku | $0.25 | $1.25 | 200k | 4k | |
| Anthropic | Claude 3 Opus | $15 | $75 | 200k | 4k | |
| Anthropic | Claude 3 Sonnet | $3 | $15 | 200k | 4k | |
| DeepSeek | DeepSeek-R1 | $0.55 | $2.19 | 64k | 8k | |
| DeepSeek | DeepSeek-V3 | $0.27 | $1.1 | 64k | 8k | |
| Gemini 1.5 Pro | $1.25 | $5 | 1m | 8k | ||
| gemini-2.0-flash-exp | $0.15 | $0.15 | 1M | 1M | ||
| Meta | Llama 3.2 90B | $0.00204 | $0.00204 | 128k | 2k | |
| Mistral | Mistral Large 24.11 | $2 | $6 | 128k | 4k | |
| OpenAI | gpt-45 | $75 | $150 | 128k | 16k | ≈200B |
| OpenAI | gpt-4o | $2.5 | $10 | 128k | 16k | ≈200B |
| OpenAI | gpt-4o-mini | $0.15 | $0.6 | 128k | 16k | ≈8B |
| OpenAI | o1 | $15 | $60 | 200k | 100k | |
| OpenAI | o1-mini | $1.1 | $4.4 | 128k | 65k | ≈100B |
| OpenAI | o3-mini | $1.1 | $4.4 | 200k | 100k |
Cost Notes:
- Input and output prices are normalized at per million tokens
- Batch runs with OpenAI are 50% less expensive
- Cached input prompts are half cost as well
Size Notes:
- Phi-3-7B, a Small Language Model (SLM) with 7 billion parameters [Abdin et al., 2024]
- Claude 3.5 Sonnet (2024-10-22), the latest model (≈175B parameters) from the Claude 3.5 family offering state-of-the-art performance across several coding, vision, and reasoning tasks [Anthropic, 2024].
- Gemini 2.0 Flash: the latest/most advanced Gemini model [Google, 2024]. Other Google models such as Med-PaLM models (540B) [Singhal et al., 2023], designed for medical purposes, were not publicly available.
- ChatGPT (≈175B) [OpenAI, 2023a] and GPT-4 (≈1.76T), a “high-intelligence” model [OpenAI, 2023b].
- GPT-4o (≈200B) providing “GPT-4-level intelligence but faster” [OpenAI, 2024a] and the GPT-4o-mini (gpt-4o-2024-05-13) small model (≈8B parameters) for focused tasks [OpenAI, 2024b].
- The latest o1-mini (o1-mini-2024-09-12) model (≈100B) [OpenAI, 2024c], and o1-preview (o1-preview-2024- 09-12) model (≈300B) with “new AI capabilities” for complex reasoning tasks [OpenAI, 2024d].
The exact numbers of parameters of several LLMs (e.g., GPT, Gemini 2.0 Flash) have not been publicly disclosed yet. Most numbers of parameters are estimate reported to provide more context for understanding the models’ performance.
From Zhang et al. (2024)