StackProofStackProof

LLM API pricing by model (August 2026)

Ranked by cost per task on a fixed profile: 1,500 input / 500 output tokens per call, 3 calls per task. Computed from the open dataset.

#ModelProviderTierInput $/1MOutput $/1MOutput share$/task$/1M tasks
1Mistral SmallMistralvalue$0.15$0.657%$0.00158$1,575
2DeepSeek-V4-FlashDeepSeekvalue$0.44$1.3250%$0.00396$3,960
3Gemini 2.5 FlashGooglefast$0.3$2.574%$0.00510$5,100
4Grok 4.3xAImid$1.25$2.540%$0.00938$9,375
5GPT-5.4 miniOpenAImid$0.75$4.567%$0.01013$10,125
6Claude Haiku 4.5Anthropicfast$1$563%$0.01200$12,000
7Claude Sonnet 5Anthropicmid$2$1063%$0.02400$24,000
8Gemini 3.1 ProGooglefrontier$2$1267%$0.02700$27,000
9Claude Opus 4.8Anthropicfrontier$5$2563%$0.06000$60,000
10GPT-5.5OpenAIfrontier$5$3067%$0.06750$67,500

The spread top to bottom is 42.9x for identical work. Before optimising anything else, check whether a cheaper model passes your evaluation cases: the cost-cutting guide ranks every lever by measured return, and the head-to-head pages put exact numbers on any two models.