Cost levers - verified
Cut your LLM API bill: prompt caching, batch and service tiers by provider
We read the prompt caching, batch and service tier docs of 15 LLM API providers on October 18, 2026. 9 take 50% off batch jobs across their batch models, and 4 discount select models only: Amazon Bedrock (50%), xAI (20%), Moonshot Kimi (40%) and Together AI (50%); DeepSeek and Z.ai document no batch API. 8 cache repeated prompts automatically, 6 cache automatically and also let you set explicit breakpoints, and only Anthropic caches nothing until you mark it. Of the 10 with per-model prices on this site, 6 typically bill a cache hit at 10% of the input price or less.
Per-model prices - Does batch stack with caching? - Cache write costs
- Batch 50% off
- 9
- Cache automatically
- 14
- Flex tier 50% off
- 5
- Batch + cache stack
- 3
Every provider, side by side
Short labels are our reading of each provider's docs; open a provider for every verbatim quote and source. The cache read discount is computed from first-party per-model prices: the median across current models, with the range when models differ. Derived means computed from a pricing table rather than stated in prose.
| Provider | Caching | Cache TTL | Cache read discounttypical, range across models | Batch discountand window | Service tiersFlex, Priority, Fast | Other levers |
|---|---|---|---|---|---|---|
| OpenAI | Automatic and explicit | 30 min (GPT-5.6+); earlier models 5 to 10 min, up to 24 hours | 90% off50% to 95%, 24 models | 50% off24h |
|
|
| Anthropic | Explicit only | 5 min, refreshed on each hit, or 1 hour | 90% off90% to 97.5%, 15 models | 50% off24h (most under 1 hour) |
|
|
| Gemini | Automatic and explicit | Explicit: 1 hour default, no bounds; implicit not stated | 90% off90% to 95%, 9 Google models | 50% off24h target, expires at 48 hours |
|
|
| Vertex AI | Automatic and explicit | Implicit: up to 24 hours; explicit: 60 min default | Implicit: 90% discount on cached tokens. Explicit: 90% discount on Gemini 2.5 or later, 75% on Gemini 2.0.From the docs; no per-model prices here | 50% off24h after start, 72 hours in queue |
|
|
| Amazon Bedrock | Automatic and explicit | 5 min default; 1 hour on Claude 4.5+; 30 min on OpenAI GPT-5.6/6.1 | 75% offsame on every model, 4 Amazon models | 50% off select models24 to 168 hour job timeout |
|
|
| Azure OpenAI | Automatic and explicit | 30 min (GPT-5.6+); earlier models 5 to 10 min, up to 24 hours | Discount on input token pricing for Standard deployments; up to 100% discount on input tokens for Provisioned deploymentsFrom the docs; no per-model prices here | 50% off24h target, jobs not expired |
|
|
| Mistral | Automatic | Not stated | 90% off89.7% to 90%, 8 Mistral AI models | 50% off24-hour period |
| None documented |
| xAI | Automatic | Not fixed: evicted under load | 84% off75% to 85%, 8 models | 20% off select models24 hours, best effort |
| None documented |
| DeepSeek | Automatic | Hours to days, cleared once unused | 97.3% off96.7% to 98%, 2 models | No batch API documented | None documented |
|
| Moonshot Kimi | Automatic | 5 min default or 1 hour, renewed on each hit | 81.6% off80% to 90%, 4 Moonshot AI models | 40% off select models12h to 7d |
| None documented |
| Z.ai | Automatic | Not specified | 81.5% off79.7% to 85.7%, 12 models | No batch API documented | None documented | None documented |
| Alibaba Qwen | Automatic and explicit | Explicit: 5 min, reset on hit; implicit: indeterminate | 90% off80% to 91.5%, 17 Alibaba Cloud (Qwen) models | 50% off24h to 336h (14d) | None documented |
|
| Groq | Automatic | 2 hours without use | 50% discount on cached input tokensFrom the docs; no per-model prices here | 50% off24h to 7d |
| None documented |
| Together AI | Automatic | Best effort, short-lived | Steep discount on supported models (per-model Cached input price)From the docs; no per-model prices here | 50% off select models24h, best effort | None documented |
|
| Fireworks AI | Automatic | Several minutes to several hours | Serverless: default 50% discount on cached prompt tokens, varies by model. Dedicated: cached tokens close to free (capacity benefit).From the docs; no per-model prices here | 50% off12, 24, 48 or 72 hours |
|
|
Per-model prices with cache and batch rates
Every model on this site with a cache read or batch price, in USD per 1M tokens from the provider's own pricing page. Cache saves is 1 minus cached input over input; batch saves compares batch to standard rates.
143 models
| Model | Input | Cached input | Batch input | Output | Batch output | Cache saveson input | Batch saves |
|---|---|---|---|---|---|---|---|
| No models match these filters. | |||||||
| Qwen3.8-Omni-FlashAlibaba Cloud (Qwen) | $0.15 | $0.016 | -not offered | $0.47 | -not offered | 89.3% | -not offered |
| Qwen3.8-FlashAlibaba Cloud (Qwen) | $0.15 | $0.016 | -not offered | $0.47 | -not offered | 89.3% | -not offered |
| Qwen3.8-27BAlibaba Cloud (Qwen) | $0.50 | $0.05 | -not offered | $3 | -not offered | 90% | -not offered |
| Qwen3.8-2.4T-A95BAlibaba Cloud (Qwen) | $2 | $0.17 | -not offered | $6 | -not offered | 91.5% | -not offered |
| Qwen3.8-MaxAlibaba Cloud (Qwen) | $2 | $0.17 | -not offered | $6 | -not offered | 91.5% | -not offered |
| Qwen3.7-FlashAlibaba Cloud (Qwen) | $0.03 | $0.003 | -not offered | $0.13 | -not offered | 90% | -not offered |
| Qwen3.7-PlusAlibaba Cloud (Qwen) | $0.40 | $0.04 | $0.20 | $1.60 | $0.80 | 90% | 50% |
| Qwen3.7-MaxAlibaba Cloud (Qwen) | $2.50 | $0.25 | -not offered | $7.50 | -not offered | 90% | -not offered |
| Qwen3.6-FlashAlibaba Cloud (Qwen) | $0.25 | $0.025 | -not offered | $1.50 | -not offered | 90% | -not offered |
| Qwen3.6-PlusAlibaba Cloud (Qwen) | $0.50 | $0.05 | -not offered | $3 | -not offered | 90% | -not offered |
| Qwen3.5-FlashAlibaba Cloud (Qwen) | $0.10 | $0.01 | -not offered | $0.40 | -not offered | 90% | -not offered |
| Qwen3.5-PlusAlibaba Cloud (Qwen) | $0.40 | $0.04 | -not offered | $2.40 | -not offered | 90% | -not offered |
| Qwen3-VL-PlusAlibaba Cloud (Qwen) | $0.20 | $0.02 | -not offered | $1.60 | -not offered | 90% | -not offered |
| Qwen-FlashAlibaba Cloud (Qwen) | $0.05 | $0.005 | $0.025 | $0.40 | $0.20 | 90% | 50% |
| Qwen3-Coder-FlashAlibaba Cloud (Qwen) | $0.30 | $0.03 | -not offered | $1.50 | -not offered | 90% | -not offered |
| Qwen-MaxAlibaba Cloud (Qwen) | $1.60 | $0.32 | $0.80 | $6.40 | $3.20 | 80% | 50% |
| Qwen-PlusAlibaba Cloud (Qwen) | $0.40 | $0.04 | $0.20 | $1.20 | $0.60 | 90% | 50% |
| Amazon Nova 2 LiteAmazon | $0.30 | $0.075 | $0.15 | $2.50 | $1.25 | 75% | 50% |
| Amazon Nova PremierAmazon | $2.50 | -not offered | $1.25 | $12.50 | $6.25 | -not offered | 50% |
| Amazon Nova LiteAmazon | $0.06 | $0.015 | $0.03 | $0.24 | $0.12 | 75% | 50% |
| Amazon Nova MicroAmazon | $0.035 | $0.00875 | $0.0175 | $0.14 | $0.07 | 75% | 50% |
| Amazon Nova ProAmazon | $0.80 | $0.20 | $0.40 | $3.20 | $1.60 | 75% | 50% |
| Claude Haiku 5.5Anthropic | $0.10 | $0.01 | $0.05 | $0.50 | $0.25 | 90% | 50% |
| Claude Sonnet 5.5Anthropic | $2 | $0.10 | $1 | $10 | $5 | 95% | 50% |
| Claude Opus 5.5Anthropic | $4 | $0.20 | $2 | $20 | $10 | 95% | 50% |
| Claude Fable 5.1Anthropic | $10 | $0.25 | $5 | $50 | $25 | 97.5% | 50% |
| Claude Mythos 5.1Anthropic | $10 | $0.25 | $5 | $50 | $25 | 97.5% | 50% |
| Claude Opus 5Anthropic | $5 | $0.50 | $2.50 | $25 | $12.50 | 90% | 50% |
| Claude Sonnet 5Anthropic | $2 | $0.20 | $1 | $10 | $5 | 90% | 50% |
| Claude Fable 5Anthropic | $10 | $1 | $5 | $50 | $25 | 90% | 50% |
| Claude Mythos 5Anthropic | $10 | $1 | $5 | $50 | $25 | 90% | 50% |
| Claude Opus 4.8Anthropic | $5 | $0.50 | $2.50 | $25 | $12.50 | 90% | 50% |
| Claude Opus 4.7Anthropic | $5 | $0.50 | $2.50 | $25 | $12.50 | 90% | 50% |
| Claude Sonnet 4.6Anthropic | $3 | $0.30 | $1.50 | $15 | $7.50 | 90% | 50% |
| Claude Opus 4.6Anthropic | $5 | $0.50 | $2.50 | $25 | $12.50 | 90% | 50% |
| Claude Opus 4.5Anthropic | $5 | $0.50 | $2.50 | $25 | $12.50 | 90% | 50% |
| Claude Haiku 4.5Anthropic | $1 | $0.10 | $0.50 | $5 | $2.50 | 90% | 50% |
| Claude Sonnet 4.5Anthropic | $3 | $0.30 | $1.50 | $15 | $7.50 | 90% | 50% |
| ERNIE 4.5 Turbo (128K)Baidu | $0.11 | $0.03 | -not offered | $0.45 | -not offered | 72.7% | -not offered |
| ERNIE 4.5 Turbo (32K)Baidu | $0.11 | $0.03 | -not offered | $0.45 | -not offered | 72.7% | -not offered |
| ERNIE 4.5 Turbo VLBaidu | $0.42 | $0.11 | -not offered | $1.27 | -not offered | 73.8% | -not offered |
| ERNIE 4.5 Turbo VL (32K)Baidu | $0.42 | $0.11 | -not offered | $1.27 | -not offered | 73.8% | -not offered |
| Dola Seed 2.1 TurboByteDance Seed | $0.50 | $0.10 | $0.25 | $2.50 | $1.25 | 80% | 50% |
| Seed 2.0 LiteByteDance Seed | $0.25 | $0.05 | -not offered | $2 | -not offered | 80% | -not offered |
| Seed 2.0 MiniByteDance Seed | $0.10 | $0.02 | -not offered | $0.40 | -not offered | 80% | -not offered |
| Seed 2.0 ProByteDance Seed | $0.50 | $0.10 | $0.25 | $3 | $1.50 | 80% | 50% |
| DeepSeek V4.1 FlashDeepSeek | $0.30 | $0.006 | -not offered | $1.20 | -not offered | 98% | -not offered |
| DeepSeek V4 ProDeepSeek | $1.32 | $0.044 | -not offered | $3.96 | -not offered | 96.7% | -not offered |
| Gemini 4 ArgonGoogle | $2 | $0.10 | $1 | $10 | $5 | 95% | 50% |
| Gemini 3.8 FlashGoogle | $0.75 | $0.075 | $0.375 | $3.75 | $1.875 | 90% | 50% |
| Gemini Robotics ER 2 PreviewGoogle | $1 | $0.10 | $0.50 | $5 | $2.50 | 90% | 50% |
| Gemini 3.5 Flash-LiteGoogle | $0.30 | $0.03 | $0.15 | $2.50 | $1.25 | 90% | 50% |
| Gemini 3.6 FlashGoogle | $0.75 | $0.075 | $0.375 | $3.75 | $1.875 | 90% | 50% |
| Gemini 3.1 Flash-LiteGoogle | $0.25 | $0.025 | $0.125 | $1.50 | $0.75 | 90% | 50% |
| Gemini 3.1 Pro PreviewGoogle | $2 | $0.20 | $1 | $12 | $6 | 90% | 50% |
| Gemini 3 Flash PreviewGoogle | $0.50 | $0.05 | $0.25 | $3 | $1.50 | 90% | 50% |
| Gemini 2.5 Flash-LiteGoogle | $0.10 | $0.01 | $0.05 | $0.40 | $0.20 | 90% | 50% |
| Gemini 2.5 FlashGoogle | $0.30 | $0.03 | $0.15 | $2.50 | $1.25 | 90% | 50% |
| Gemini 2.5 ProGoogle | $1.25 | $0.125 | $0.625 | $10 | $5 | 90% | 50% |
| Mercury 2.5Inception | $0.20 | $0.02 | -not offered | $0.75 | -not offered | 90% | -not offered |
| Mercury Edit 2Inception | $0.25 | $0.025 | -not offered | $0.75 | -not offered | 90% | -not offered |
| Mercury 2Inception | $0.25 | $0.025 | -not offered | $0.75 | -not offered | 90% | -not offered |
| MiniMax-M3MiniMax | $0.30 | $0.06 | -not offered | $1.20 | -not offered | 80% | -not offered |
| MiniMax-M2.7MiniMax | $0.30 | $0.06 | -not offered | $1.20 | -not offered | 80% | -not offered |
| MiniMax-M2.7-highspeedMiniMax | $0.60 | $0.06 | -not offered | $2.40 | -not offered | 90% | -not offered |
| MiniMax-M2.5MiniMax | $0.30 | $0.03 | -not offered | $1.20 | -not offered | 90% | -not offered |
| MiniMax-M2.5-highspeedMiniMax | $0.60 | $0.03 | -not offered | $2.40 | -not offered | 95% | -not offered |
| Mistral Large 4Mistral AI | $1.36 | $0.14 | $0.68 | $4.18 | $2.09 | 89.7% | 50% |
| Mistral Medium 3.5Mistral AI | $1.50 | $0.15 | $0.75 | $7.50 | $3.75 | 90% | 50% |
| Mistral Small 4Mistral AI | $0.15 | $0.015 | $0.075 | $0.60 | $0.30 | 90% | 50% |
| Ministral 3 14BMistral AI | $0.20 | $0.02 | $0.10 | $0.20 | $0.10 | 90% | 50% |
| Ministral 3 3BMistral AI | $0.10 | $0.01 | $0.05 | $0.10 | $0.05 | 90% | 50% |
| Ministral 3 8BMistral AI | $0.15 | $0.015 | $0.075 | $0.15 | $0.075 | 90% | 50% |
| Mistral Large 3Mistral AI | $0.50 | $0.05 | $0.25 | $1.50 | $0.75 | 90% | 50% |
| Codestral 25.08Mistral AI | $0.30 | $0.03 | $0.15 | $0.90 | $0.45 | 90% | 50% |
| Kimi K2.6Moonshot AI | $0.95 | $0.16 | $0.57 | $4 | $2.40 | 83.2% | 40% |
| Kimi K2.7 CodeMoonshot AI | $0.95 | $0.19 | $0.57 | $4 | $2.40 | 80% | 40% |
| Kimi K2.7 Code HighSpeedMoonshot AI | $1.90 | $0.38 | -not offered | $8 | -not offered | 80% | -not offered |
| Kimi K3Moonshot AI | $3 | $0.30 | -not offered | $15 | -not offered | 90% | -not offered |
| GPT-6.1 SolOpenAI | $2 | $0.10 | $1 | $10 | $5 | 95% | 50% |
| GPT-6 LunaOpenAI | $0.10 | $0.01 | $0.05 | $0.50 | $0.25 | 90% | 50% |
| GPT-6 SolOpenAI | $2 | $0.20 | $1 | $10 | $5 | 90% | 50% |
| GPT-6 AstraOpenAI | $10 | $1 | $5 | $50 | $25 | 90% | 50% |
| GPT-5.6 CyberOpenAI | $12.50 | $1.25 | -not offered | $75 | -not offered | 90% | -not offered |
| GPT-5.6 LunaOpenAI | $0.20 | $0.02 | $0.10 | $1.20 | $0.60 | 90% | 50% |
| GPT-5.6 SolOpenAI | $4 | $0.40 | $2 | $20 | $10 | 90% | 50% |
| GPT-5.6 TerraOpenAI | $2 | $0.20 | $1 | $12 | $6 | 90% | 50% |
| GPT-5.5OpenAI | $5 | $0.50 | $2.50 | $30 | $15 | 90% | 50% |
| GPT-5.5 ProOpenAI | $30 | -not offered | $15 | $180 | $90 | -not offered | 50% |
| GPT-5.4 MiniOpenAI | $0.75 | $0.075 | $0.375 | $4.50 | $2.25 | 90% | 50% |
| GPT-5.4 nanoOpenAI | $0.20 | $0.02 | $0.10 | $1.25 | $0.625 | 90% | 50% |
| GPT-5.4OpenAI | $2.50 | $0.25 | $1.25 | $15 | $7.50 | 90% | 50% |
| GPT-5.3-CodexOpenAI | $1.75 | $0.175 | -not offered | $14 | -not offered | 90% | -not offered |
| GPT-5.2OpenAI | $1.75 | $0.175 | $0.875 | $14 | $7 | 90% | 50% |
| GPT-5.1OpenAI | $1.25 | $0.125 | $0.625 | $10 | $5 | 90% | 50% |
| GPT-5 ProOpenAI | $15 | -not offered | $7.50 | $120 | $60 | -not offered | 50% |
| GPT-5OpenAI | $1.25 | $0.125 | $0.625 | $10 | $5 | 90% | 50% |
| GPT-5 MiniOpenAI | $0.25 | $0.025 | $0.125 | $2 | $1 | 90% | 50% |
| GPT-5 nanoOpenAI | $0.05 | $0.005 | $0.025 | $0.40 | $0.20 | 90% | 50% |
| o3-proOpenAI | $20 | -not offered | $10 | $80 | $40 | -not offered | 50% |
| o3OpenAI | $2 | $0.50 | $1 | $8 | $4 | 75% | 50% |
| o4-miniOpenAI | $1.10 | $0.275 | $0.55 | $4.40 | $2.20 | 75% | 50% |
| GPT-4.1OpenAI | $2 | $0.50 | $1 | $8 | $4 | 75% | 50% |
| GPT-4.1 MiniOpenAI | $0.40 | $0.10 | $0.20 | $1.60 | $0.80 | 75% | 50% |
| GPT-4.1 nanoOpenAI | $0.10 | $0.025 | $0.05 | $0.40 | $0.20 | 75% | 50% |
| o1-proOpenAI | $150 | -not offered | $75 | $600 | $300 | -not offered | 50% |
| o3-miniOpenAI | $1.10 | $0.55 | $0.55 | $4.40 | $2.20 | 50% | 50% |
| o1OpenAI | $15 | $7.50 | $7.50 | $60 | $30 | 50% | 50% |
| GPT-4o MiniOpenAI | $0.15 | $0.075 | $0.075 | $0.60 | $0.30 | 50% | 50% |
| GPT-4oOpenAI | $2.50 | $1.25 | $1.25 | $10 | $5 | 50% | 50% |
| GPT-4 TurboOpenAI | $10 | -not offered | $5 | $30 | $15 | -not offered | 50% |
| Chat LatestOpenAI | $5 | $0.50 | -not offered | $30 | -not offered | 90% | -not offered |
| GPT-3.5 TurboOpenAI | $0.50 | -not offered | $0.25 | $1.50 | $0.75 | -not offered | 50% |
| GPT-4OpenAI | $30 | -not offered | $15 | $60 | $30 | -not offered | 50% |
| Step 3.7 FlashStepFun | $0.20 | $0.04 | -not offered | $1.15 | -not offered | 80% | -not offered |
| Step 3.5 FlashStepFun | $0.10 | $0.02 | -not offered | $0.30 | -not offered | 80% | -not offered |
| Step 3.5 Flash 2603StepFun | $0.10 | $0.02 | -not offered | $0.30 | -not offered | 80% | -not offered |
| Step 5 PreviewStepFun | $1 | $0.10 | -not offered | $2.70 | -not offered | 90% | -not offered |
| Step-1o Turbo VisionStepFun | $0.36 | $0.07 | -not offered | $1.15 | -not offered | 80.6% | -not offered |
| Hy4 previewTencent | $0.834 | $0.042 | -not offered | $2.501 | -not offered | 95% | -not offered |
| Hy3Tencent | $0.132 | $0.033 | -not offered | $0.528 | -not offered | 75% | -not offered |
| Grok 4.7xAI | $2 | $0.50 | -not offered | $6 | -not offered | 75% | -not offered |
| Grok 4.6xAI | $2 | $0.50 | -not offered | $6 | -not offered | 75% | -not offered |
| Grok 4.5xAI | $2 | $0.30 | -not offered | $6 | -not offered | 85% | -not offered |
| Grok Build 0.1xAI | $1 | $0.20 | -not offered | $2 | -not offered | 80% | -not offered |
| Grok 4.20xAI | $1.25 | $0.20 | $1 | $2.50 | $2 | 84% | 20% |
| Grok 4.20 (Non-Reasoning)xAI | $1.25 | $0.20 | $1 | $2.50 | $2 | 84% | 20% |
| Grok 4.20 Multi-Agent BetaxAI | $1.25 | $0.20 | $1 | $2.50 | $2 | 84% | 20% |
| Grok 4.3xAI | $1.25 | $0.20 | $1 | $2.50 | $2 | 84% | 20% |
| MiMo-V2.6-FlashXiaomi MiMo | $0.14 | $0.0028 | $0.07 | $0.28 | $0.14 | 98% | 50% |
| MiMo-V2.6-ProXiaomi MiMo | $0.435 | $0.0036 | $0.2175 | $0.87 | $0.435 | 99.2% | 50% |
| GLM-5.3-FlashZ.ai | $0.15 | $0.03 | -not offered | $0.50 | -not offered | 80% | -not offered |
| GLM-5.3Z.ai | $1.40 | $0.26 | -not offered | $4.40 | -not offered | 81.4% | -not offered |
| GLM-5.2Z.ai | $1.40 | $0.26 | -not offered | $4.40 | -not offered | 81.4% | -not offered |
| GLM-5.1Z.ai | $1.40 | $0.26 | -not offered | $4.40 | -not offered | 81.4% | -not offered |
| GLM-5Z.ai | $1 | $0.20 | -not offered | $3.20 | -not offered | 80% | -not offered |
| GLM-4.7Z.ai | $0.60 | $0.11 | -not offered | $2.20 | -not offered | 81.7% | -not offered |
| GLM-4.6VZ.ai | $0.30 | $0.05 | -not offered | $0.90 | -not offered | 83.3% | -not offered |
| GLM-4.6Z.ai | $0.60 | $0.11 | -not offered | $2.20 | -not offered | 81.7% | -not offered |
| GLM-4.5Z.ai | $0.60 | $0.11 | -not offered | $2.20 | -not offered | 81.7% | -not offered |
| GLM-4.5-AirZ.ai | $0.20 | $0.03 | -not offered | $1.10 | -not offered | 85% | -not offered |
| GLM-4.7-FlashXZ.ai | $0.07 | $0.01 | -not offered | $0.40 | -not offered | 85.7% | -not offered |
| GLM-5.3-FlashXZ.ai | $0.37 | $0.075 | -not offered | $1.25 | -not offered | 79.7% | -not offered |
Does batch stack with caching?
Only from what each provider's own pages say. Not stated means the pages checked do not answer it, not that the answer is no.
Yes: the batch discount and the cache discount combine
- Anthropic
“Batch API and prompt caching discounts can be combined.”
platform.claude.com - xAI
“The batch discount applies to all token types”
Sources: docs.x.ai/developers/advanced-api-usage/batch-api, docs.x.ai/developers/pricing
- Fireworks AI
“Batch jobs automatically use prompt caching for additional 50% cost savings on cached tokens.”
Sources: docs.fireworks.ai/guides/batch-inference, docs.fireworks.ai/serverless/pricing
No: you get one discount or the other, not both
- Vertex AI
“However, the discounts for cache and batch don't stack. The 90% cache hit discount takes precedence over the batch discount.”
docs.cloud.google.com - Alibaba Qwen
“It also does not support features such as context cache.”
alibabacloud.com - Groq
“Note: The batch discount does not stack with prompt caching discounts. All batch tokens are billed at the 50% batch rate regardless of cache status.”
console.groq.com
Cache hits inside a batch are billed at the standard caching rate
- Gemini
“If a request in your batch results in a cache hit, you pay the standard context caching rates.”
ai.google.dev
Not stated: the provider's pages do not say
OpenAI, Amazon Bedrock, Azure OpenAI, Mistral, Moonshot Kimi, Together AI
Minimum cacheable prompt
A prompt shorter than the minimum is never cached, so a short system prompt earns no discount. 9 of 15 providers state a minimum.
| Provider | Minimum cacheable prompt |
|---|---|
| OpenAI | GPT-5.6 and later: 1,024 visible input tokens. Earlier models: varies by request settings. |
| Anthropic | 512 tokens: Fable 5.1, Mythos 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Fable 5, Mythos 5, Haiku 5.5. 1,024: Opus 4.8, Sonnet 5, Sonnet 4.6, Sonnet 4.5, Opus 4.1, Opus 4, Sonnet 4. 2,048: Mythos Preview, Opus 4.7, Haiku 3.5. 4,096: Opus 4.6, Opus 4.5, Haiku 4.5. |
| Gemini | Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.6 Flash, Gemini 3.1 Pro Preview: 4,096. Gemini 2.5 Flash and 2.5 Pro: 2,048. |
| Vertex AI | Gemini 3 (with exceptions) and Gemini 4 models: 4,096 tokens. Gemini 3.0 Flash Preview, 3.1 Pro Preview, 3.7 Flash, and 3.8 Flash (implicit caching only): 6,144 tokens. Gemini 2 family: 2,048 tokens. |
| Amazon Bedrock | Per checkpoint, cumulative prefix: ranges from 512 to 4,096 by model (Claude 5.x family 512; Opus 4.8 / Sonnet 5 / Sonnet 4.x / Claude 3.x 1,024; Opus 4.5-4.7 and Haiku 4.5 4,096; Nova 1K; OpenAI GPT-5.6/6.1 1,024) |
| Azure OpenAI | 1,024 tokens, first 1,024 identical |
| Mistral | 64 tokens (cache blocks of 64 tokens; prompts under 64 tokens get no hits) |
| xAI | Not stated |
| DeepSeek | Not stated |
| Moonshot Kimi | Not stated; cache stored in blocks, and a portion smaller than one full block is a cache miss |
| Z.ai | Not stated |
| Alibaba Qwen | 1,024 tokens for both modes (512 for Zhipu-deployed GLM models on implicit cache) |
| Groq | 128 to 1024 tokens depending on the model |
| Together AI | Not stated |
| Fireworks AI | Not stated |
Cache writes cost extra
A cache hit is cheap, but the first request that writes the cache can cost more than plain input. In the per-model prices on this site, OpenAI, Anthropic and Alibaba Qwen list writes above the input price on at least one model; Gemini, Vertex AI and Z.ai also bill cache storage.
| Provider | Write price vs inputcomputed from per-model prices | What the docs say | Storage |
|---|---|---|---|
| OpenAI | 1.25x | GPT-5.6 and later: cache writes 1.25x the uncached input rate. GPT-5.5 and earlier: no additional cache-write charge. | Not stated |
| Anthropic | 1.25x1-hour write 2x | 5-minute writes 1.25x base input; 1-hour writes 2x base input | Not stated |
| Gemini | No write price listed | Not stated | Explicit caches: billed per hour of TTL on the cached token count (per-model storage price, e.g. $1.00 / 1M tokens per hour on Gemini 3.x Flash models after promo; $4.50 / 1M tokens per hour on Gemini 3.1 Pro Preview). Implicit: no storage line documented. |
| Vertex AI | No write price listed | Cache creation billed at the standard input token price (no premium) | Explicit caching only: storage costs based on how long caches are stored. No storage cost for implicit caching. |
| Amazon Bedrock | No write price listed | Model-dependent: writes can be billed above standard input; OpenAI models on Bedrock bill writes at 1.25x uncached input | Not stated |
| Azure OpenAI | No write price listed | Models before GPT-5.6: no extra charge to write. GPT-5.6 and later: cache writes can incur charges in addition to discounted reads. | Not stated |
| Mistral | No write price listed | No separate cache-write charge described; uncached tokens billed at the standard input price | Not stated |
| xAI | No write price listed | Not stated | Not stated |
| DeepSeek | No write price listed | Not stated | Not stated |
| Moonshot Kimi | 1x1-hour write 2x | kimi-k3: cache writes billed separately, 5m write at the input price, 1h write at 2x; K2 series (kimi-k2.7, k2.7-highspeed, k2.6) have no cache-write chargeDerived from the pricing table | Not stated |
| Z.ai | No write price listed | No cache-write line item; new content billed at standard prices | Cached Input Storage listed as 'Limited-time Free' on the pricing table |
| Alibaba Qwen | 1.25x to 1.33x | Explicit: typically 125% of standard input (only the incremental portion when extending an existing cache). Implicit: 100% of standard input (no premium). | Not stated |
| Groq | No write price listed | No additional fees | Not stated |
| Together AI | No write price listed | Not stated | Not stated |
| Fireworks AI | No write price listed | Not stated | Not stated |
FAQ
Which LLM APIs offer a batch discount?
Of the 15 providers checked on October 18, 2026, 13 document a batch API. OpenAI, Anthropic, Gemini, Vertex AI, Azure OpenAI, Mistral, Alibaba Qwen, Groq and Fireworks AI take 50% off. Amazon Bedrock (50%), xAI (20%), Moonshot Kimi (40%) and Together AI (50%) discount listed models only. DeepSeek and Z.ai document no batch API.
Does the batch discount stack with prompt caching?
They stack at Anthropic, xAI and Fireworks AI. They do not at Vertex AI, Alibaba Qwen and Groq. Gemini bills a cache hit inside a batch at the standard caching rate. OpenAI, Amazon Bedrock, Azure OpenAI, Mistral, Moonshot Kimi and Together AI do not say.
Which LLM APIs cache prompts automatically?
14 of 15 cache repeated prompt prefixes with no code change: Mistral, xAI, DeepSeek, Moonshot Kimi, Z.ai, Groq, Together AI, Fireworks AI, OpenAI, Gemini, Vertex AI, Amazon Bedrock, Azure OpenAI and Alibaba Qwen. OpenAI, Gemini, Vertex AI, Amazon Bedrock, Azure OpenAI and Alibaba Qwen also offer explicit breakpoints or cache objects. Anthropic caches only what you mark with cache_control.
How much does a prompt cache hit save?
Computed from first-party per-model prices (typical saving on input, with the range across current models): OpenAI 90% (50% to 95%); Anthropic 90% (90% to 97.5%); Gemini 90% (90% to 95%); Amazon Bedrock 75%; Mistral 90% (89.7% to 90%); xAI 84% (75% to 85%); DeepSeek 97.3% (96.7% to 98%); Moonshot Kimi 81.6% (80% to 90%); Z.ai 81.5% (79.7% to 85.7%); Alibaba Qwen 90% (80% to 91.5%). Several providers also charge more than the input price to write the cache.
Verified from each provider's caching, batch, service tier and pricing docs; every quote was matched against the live page. Per-model prices come from the model registry, each dated by its own verification day. Discounts change without notice; check the linked source before you commit a budget to one.
Newsletter
Know when a provider changes its prices
One email per week. New model releases, pricing changes and what they mean for what you build.
Keep exploring
Model API pricing and specs
First-party prices per 1M tokens and sources for every model.
Free LLM API tiers
Which APIs are free to start, with free limits and data terms.
AI rate limits and usage limits
Paid-tier limits per provider, copied verbatim from the docs.
AI cost calculator
Estimate what a workload costs across models.