AI model API pricing

All models by price

Individual model pricing guides

Detailed breakdown per model: token costs, context limits, caching structures, use-case recommendations, and cost scenarios.

Pricing articles and analysis

Deeper reading on cost strategy, provider comparisons, and how to reduce your API bill.

• Lowest-cost APIs tracked by benchr → — DeepSeek, Llama, and where the quality/cost line breaks down.
• Price per use case → — The cheapest model for chat, coding, RAG, agents, and classification.
• How to reduce token usage → — Caching, batching, and routing that cut bills without switching models.
• DeepSeek vs OpenAI pricing → — What the cost gap means on real workloads.
• Claude API pricing guide → — Opus 5, Sonnet 5, Haiku 4.5 and Fable 5.1, plus the older tiers, with prompt-caching discounts explained.
• OpenAI API pricing guide → — GPT-6 Astra, the GPT-5.6 family and the older GPT-5 tiers, with batch and caching discounts.
• AI model pricing comparison → — All ranked models side by side with notes on what each price tier buys.

How to read these pricing tables

API providers bill in tokens — roughly 750 words per 1,000 tokens for English text. Prices are listed per million tokens because per-token rates are too small to read ($0.000001250 is clearer as $1.25/1M). Most providers charge separately for input (text you send) and output (text the model generates).

Output tokens typically cost 4–12× more than input tokens because they require a separate forward pass per token. This means the input-to-output ratio of your workload matters enormously. A summarization task (high input, low output) has a very different cost profile than a story-generation task (low input, high output).

Three cost strategies for 2026

Route by task complexity. Use cheap small models (GPT-5.6 Luna at $0.20 input per 1M, DeepSeek-V4.1-Flash at $0.30 peak or $0.15 off-peak, Gemini 3.5 Flash-Lite at $0.30) for classification, routing, and simple extraction. Reserve expensive flagships for tasks that require their capability. Most pipelines can route 70–80% of calls to cheaper models.

Maximize prompt caching. If your system prompt or document context repeats across calls, caching cuts that portion of your input cost by 90% (Anthropic, Google) or ~99% (DeepSeek). This single optimization often reduces monthly bills more than switching models entirely.

Use Batch API for offline workloads. OpenAI, Anthropic, and others offer 50% discounts for asynchronous batch jobs — dataset evaluation, classification runs, bulk translation. If your workload can tolerate a 24-hour return window, batch pricing cuts your bill in half.

Use the Cost Calculator to model your specific token mix, or the Cheapest API leaderboard to find the lowest-cost model that meets your quality bar.