Which AI model costs less for your workload? October 17, 2026 · Verified first-party API Standard rates; change the token mix and see the ordering change. This is not a quality benchmark . Some rates expire soon.
Input tokens per call Output tokens per call Calls Cache-hit share of input (%)
Model / tariff Uncached in / out Workload total Relative billed cost (log scale)
Price rows are not quality-adjusted: the same token count does not imply the same result. Input caching must be a genuine provider cache hit; the model's context limit and cached input rules must be verified. OpenAI prices refer to short context. Haiku 5.5's listed rate applies to prompts up to 100K. Gemini Argon rises Dec 1, Gemini 3.8 Flash rises Jan 1, and DeepSeek depends on UTC peak hours. Cache write and tool fees are omitted. Change the rates after the effective dates.
Sources: OpenAI · Anthropic · DeepSeek · Google · UsageBox task-cost calculator