- LLM costs dropped from $30/M tokens in 2023 to under $1/M for comparable capability by 2025.
- Gemini 2.5 Flash Lite ($0.075/$0.30) and GPT-4o Mini ($0.15/$0.60) offer best value for most tasks.
- Organizations spending $5,000+ monthly on API tokens find self-hosting open-weight models economically attractive.
- Semantic caching and request batching can reduce API calls by 30-60% without affecting model quality.
- Llama 4 Scout for high-volume summarization, Claude 3.5 Haiku for customer-facing applications.
The New Economics of LLMs
The cost of AI has collapsed. In 2023, GPTGPTGenerative Pre-trained Transformer — the model architecture and family name behind OpenAI's most famous models, from GPT-2 to GPT-5.Learn more →-4 cost $30 per million input tokens. By 2025, models with comparable capability cost under $1/M. The emergence of efficient small models, open-weightOpen-WeightA model whose trained weights are publicly available for download, allowing anyone to run, fine-tune, or build on top of it — distinct from fully open-source (which also includes training code and data).Learn more → alternatives, and fierce competition among providers has created an extraordinary era for budget-conscious builders.
Understanding cost requires looking at both input and output tokenTokenThe basic unit of text that an LLM processes — roughly corresponding to a word or word-piece. Models read input and produce output in tokens, which is also how API usage is measured and billed.Learn more → prices, as well as quality per dollar. A $0.10/M model that requires ten times more tokens to accomplish a task is not cheaper than a $0.50/M model that does it in two. Always benchmarkBenchmarkA standardized test or set of tasks used to evaluate and compare the capabilities of different AI models on a common scale.Learn more → cost-per-task, not cost-per-token.
Best Value Models
Gemini 2.5 Flash Lite at $0.075/$0.30 per million tokens is the cheapest capable frontier-adjacent model available. For classification, extraction, summarization, and simple Q&A at scale, it is hard to beat. GPT-4o Mini at $0.15/$0.60 remains competitive, with strong instruction-following and broad task coverage at a very low price.
Claude 3.5 Haiku ($0.80/$4.00) is slightly pricier but offers Anthropic's instruction-following quality in a fast, compact package. It is ideal for customer-facing applications where response quality has real business impact. Llama 4 Scout via API costs just $0.05/$0.17 and offers a remarkable 10M context windowContext WindowThe maximum amount of text (measured in tokens) that a model can read and reason over in a single interaction, including your prompt, any documents, and previous conversation history.Learn more →.
Self-Hosting for Maximum Savings
For high enough volumes, self-hosting open-weight models on your own infrastructure eliminates per-token costs entirely. Llama 4 Maverick, Mistral Small, and Qwen 2.5 72B are excellent candidates. The break-even point depends on your GPU costs, but organizations spending $5,000 or more per month on API tokens generally find self-hosting economically attractive.
QuantizationQuantizationA compression technique that reduces model size and inference cost by storing weights in lower-precision numerical formats, trading a small amount of accuracy for much lower memory and compute requirements.Learn more → enables running competitive 7B to 13B Foundation Models on consumer hardware with minimal quality loss. For development, testing, and lower-stakes production tasks, running Mistral 7B locally costs essentially nothing beyond electricity.
Budget Recommendations by Use Case
Simple classification and extraction: Gemini 2.5 Flash Lite or GPT-4o Mini. High-volume summarization: Llama 4 Scout via API. Customer-facing chat: Claude 3.5 Haiku or GPT-4o Mini. Development and testing: Mistral Small or a locally-run Llama variant. Cost-optimized RAGRAGRetrieval-Augmented Generation — a technique that enhances LLM responses by first retrieving relevant documents from an external knowledge base and including them in the prompt.Learn more →: embed with text-embeddingEmbeddingA numerical vector representation of text, code, or images that captures semantic meaning — similar items have similar vectors, enabling search, clustering, and retrieval.Learn more →-3-small, generate with Claude Haiku or GPT-4o Mini.
Implement semantic caching and request batching before adding model downgrades to your cost reduction strategy. Tools like Semantic Router and LLMLLMLarge Language Model — a neural network trained on vast amounts of text data that can understand and generate human-quality language across a wide range of tasks.Learn more →-cache can reduce API calls by 30 to 60 percent for repeated or similar queries, saving money without touching model quality.