AI 模型定价目录

GPT-5.6 flagship for complex professional and agentic work. The gpt-5.6 alias routes to Sol. Explicit cache writes cost 1.25x uncached input; requests above 272K input tokens use long-context pricing.

Balanced GPT-5.6 production model at OpenAI's reduced August 2026 price. Explicit cache writes cost 1.25x uncached input; requests above 272K input tokens use long-context pricing.

High-volume GPT-5.6 routing model at OpenAI's reduced August 2026 price, with the same 1.05M context window and 128K maximum output as Sol and Terra.

Previous OpenAI frontier model for complex reasoning, coding, and professional work. GPT-5.6 Sol is the current flagship; prompts above 272K input tokens use higher long-context pricing.

Higher-compute GPT-5.5 variant for the hardest professional tasks. No cached-input discount is listed on the official model page.

Legacy OpenAI model for coding and professional work. GPT-5.6 Terra now uses the same standard $2.50/$15 price point.

Highest reasoning capability for elite tasks.

Standard reasoning model.

Current lower-cost GPT-5.4 production model.

Current high-volume GPT-5.4 routing model.

Current dedicated Codex API model for long-horizon agentic coding.

Claude model for fiction and long-form creative work. Global API access was restored on 2026-07-01. The cache-write price shown is for a 5-minute TTL; 1-hour writes cost $20/M.

Current Claude Opus flagship for complex coding, agent, and knowledge-work tasks. The cache-write price shown is for a 5-minute TTL; 1-hour writes cost $10/M.

Introductory pricing through 2026-08-31. On 2026-09-01, standard pricing becomes $3/M input, $0.30/M cache reads, $3.75/M 5-minute cache writes, and $15/M output.

Previous-generation Opus model. Full 1M context remains available at standard pricing.

Previous-generation Opus model. Full 1M context is available at standard pricing.

Previous-generation balanced Claude model. Full 1M context remains available at standard pricing.

Fastest Claude model. 200K context, 64K max output.

Current GA Gemini Flash workhorse for coding, multimodal reasoning, and multi-step agents. The 2026 introductory rate also applies to Batch and Flex at half the listed synchronous rate.

Previous GA Gemini Flash model. Google applies the same introductory 2026 price as Gemini 3.7 Flash; the standard rate returns on 2027-01-01.

Current GA low-cost Gemini model for high-throughput multimodal workloads.

Stable Gemini 3.5 Flash model for agentic loops, coding cycles, long-horizon tasks, search grounding, Batch API, Flex, context caching, and multimodal inputs.

Preview Gemini Pro model. Inputs above 200K tokens use higher long-context pricing.

Fastest Gemini model. Optimized for high-throughput, multimodal tasks.

Long context >200K: $2.50 input, $15.00 output per 1M.

Current xAI frontier reasoning model with low, medium, high, and xhigh effort. xAI publishes no text-output limit. Prompts at or above 200K tokens bill all tokens at long-context rates; Priority Processing costs 2x standard rates.

Previous xAI frontier model for coding, agentic tasks, and knowledge work. xAI does not publish a separate maximum-output limit; prompts at or above 200K tokens use long-context pricing.

Lower-cost general-purpose xAI API model with configurable reasoning. xAI does not publish a separate maximum-output limit; requests above 200K tokens use long-context pricing.

Pinned Grok 4.20 reasoning endpoint. xAI does not publish a separate maximum-output limit; requests above 200K tokens use long-context pricing.

Pinned Grok 4.20 endpoint without reasoning tokens. xAI does not publish a separate maximum-output limit; requests above 200K tokens use long-context pricing.

Beta multi-agent endpoint for parallel research with 4-agent and 16-agent configurations. xAI does not publish a separate maximum-output limit.

Public-beta xAI model optimized for coding and software-building workflows. Requests above 200K tokens use long-context pricing; xAI does not publish a separate maximum-output limit.

Current balanced Mistral production model.

Current low-cost Mistral model for high-volume production routing. Mistral lists a 90% discount for cached input tokens.

Serves DeepSeek-V4-Flash-0731. New time-of-day pricing took effect on 2026-08-16 at 16:00 UTC. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; off-peak rates are 50% lower. The retired deepseek-chat and deepseek-reasoner aliases were discontinued on 2026-07-24.

Serves DeepSeek-V4-Pro-0813. New time-of-day pricing took effect on 2026-08-16 at 16:00 UTC. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; off-peak rates are 50% lower.

Open-sourced under MIT. New pay-as-you-go pricing took effect on 2026-05-27 00:00 Beijing time. Domestic pricing is ¥3.00/M cache-miss input, ¥0.025/M cached input, and ¥6.00/M output. Cache writing is currently free for a limited time.

Native full-modal model with text, image, video, and audio understanding. New pay-as-you-go pricing took effect on 2026-05-27 00:00 Beijing time. Domestic pricing is ¥1.00/M cache-miss input, ¥0.02/M cached input, and ¥2.00/M output.

Latest MiniMax frontier coding and agent model. Supports Adaptive Thinking, tool use, and native image/video input. Inputs above 512K tokens use long-context pricing and currently require limited-access availability.

Latest GLM model with published pay-as-you-go USD API pricing. GLM-5.3 is newer and available in Coding Plan, but its standard API access and token pricing are still coming soon.

Previous GLM flagship for long-horizon coding and agent tasks. It remains available, while GLM-5.2 adds a 1M-token context window.

Current faster GLM model optimized for tool use and agent workflows.

Current multimodal GLM agent model for text, image, and video input.

China API pricing in CNY. Current K2.5 multimodal model with text, image, and video input.

Global API pricing in USD for the general Kimi multimodal model.

Global API pricing in USD. Current Kimi coding model with thinking-only operation.

Always-reasoning K3 API with automatic context caching and max reasoning effort. Input plus output cannot exceed 1,048,576 tokens; the full model weights are available.

Current Qwen3.7 Plus alias promotional price in China, listed in CNY. List price is ¥2/¥8 per million tokens below 256K and ¥6/¥24 above 256K; Alibaba lists no promotion end date.

Current Qwen3.7 Max alias promotional price in China, listed in CNY. List price is ¥12/M input and ¥36/M output; Alibaba lists no promotion end date.

Current Cohere flagship for agentic, multilingual, RAG, and enterprise workloads.

Enterprise RAG-optimized with citation grounding.

Current Jamba Large API alias with a hybrid SSM-Transformer architecture and 256K context. The stable site slug is retained for existing links.

Current Jamba Mini API alias with 256K context. The stable site slug is retained for existing links.

Current Amazon Nova 2 model on Bedrock for multimodal reasoning, tool use, and long-context workloads.

Via AWS Bedrock. 300K context.

Text-only. Lowest cost option on Bedrock.