Prompt caching AI models: send the same prefix twice, pay once

Prices from official vendor price lists · October 2026
147 models that charge a fraction of the input rate for repeated prompt prefixes — often the largest saving available. Output prices run from $0.059 to $180 per 1M tokens across the 142 models with a paid per-token price; 71 have a free tier and 71 ship open weights.
Prompt caching
Services found: 147
#AI modelCapabilitiesAPI $/1M (in / out)Context windowPrice / moOutput $/1MOverviewVisit site
141

Aion 3.5 Mini

Aion Labs
Text & chatReasoningPrompt caching
$0.7 / $1.4256K tokensAPI only1.4View
142

Solar Mini 4

Upstage
Text & chatReasoningAdjustable thinking
$0.1 / $0.4512K tokensAPI only0.4View
143

Perceptron Mk1.5

Perceptron
Text & chatVisionAudio
$0.15 / $1.536.9K tokensAPI only1.5View
144

Ling 3.0 Flash VL

InclusionAI
Text & chatVisionVideo
$0.02 / $0.059256K tokensAPI only0.059View
145

Pareto

Unbiased
Text & chatVisionPrompt caching
$0.8 / $3.2API only3.2View
146

Schematron V2 Turbo

Inference.net
Text & chatStructured JSON outputPrompt caching
$0.03 / $0.15125K tokensAPI only0.15View
147

Schematron V2 Small

Inference.net
Text & chatStructured JSON outputPrompt caching
$0.05 / $0.23125K tokensAPI only0.23View

Browse by need

Reasoning models169 modelsCoding assistants17 modelsVision and images156 modelsLong context 1M+82 modelsFree tier available80 modelsOpen weights88 models

Prompt caching stores an identical prompt prefix between calls, so repeated context is charged at a fraction of the base input rate - typically 10%, and as little as 2.5% on Claude Fable 5.1. If your application sends the same lengthy system prompt with every request, this is usually the biggest saving available to you.

147 models support it. The saving applies to input only, so the effect is largest where prompts are long and answers are short - support bots with detailed instructions, agents with big tool catalogues, RAG systems that resend the same corpus. Cache entries expire, so the economics favour high request frequency against a stable prefix.

Input onlyThe discount never touches output, so it favours long prompts and short answers.
Prefix must match exactlyOne changed character at the start invalidates the whole cached block.
Frequency beats sizeA modest prompt reused constantly saves more than a huge one called twice a day.

Models with this capability

The 147 models below all support this capability. Prices are input and output per 1M tokens from the vendor official price list; for dirhams multiply by 3.6725. Tap a card to open the model page.

Claude Fable 5.1
#1
Anthropic
in $10out $501M tokens
GPT-6 Astra
#2
OpenAI
in $10out $501.05M tokens
GPT-6.1 Sol
#3
OpenAI
in $2out $101.05M tokens
GPT-6 Sol
#4
OpenAI
in $2out $101.05M tokens
GPT-6 Luna
#5
OpenAI
in $0.1out $0.51.05M tokens
Claude Opus 5.5
#6
Anthropic
in $4out $201M tokens
Claude Sonnet 5.5
#7
Anthropic
in $2out $101M tokens
Grok 4.7
#8
xAI
in $2out $6500K tokens

How to choose

  • Put everything stable at the very front: system prompt, tool definitions, reference documents.
  • Keep user-specific content after the cached prefix.
  • Check the cache lifetime — short TTLs punish low-traffic applications.
  • Account for write cost: some vendors charge a premium to create the entry.

Common mistakes

Interleaving variables into the prefixA timestamp at the top of the prompt destroys the cache on every call.
Assuming it helps everywhereShort prompts gain almost nothing.

Who this suits

Agent platformsLarge tool catalogues resent with every step.
RAG systemsThe same corpus queried repeatedly by many users.
Support botsDetailed instructions in front of every short question.

Related filters

Batch API140 models with half-price async processing
Long context 1M+82 models holding 1M tokens or more
Agentic workflows5 models built for autonomous multi-step work
Tool calling166 models that call your functions
All 239 modelsThe full rating with filters and prices

Frequently asked questions

How much does caching save?

Cache reads typically cost 10% of base input, and 2.5% on Claude Fable 5.1. The saving applies to input only.

What has to stay identical?

The prefix — system prompt, tool definitions, reference documents. Anything after the cached portion can vary freely.

When is caching pointless?

Short prompts, or requests spaced far enough apart that the cache expires between them.

How much can I realistically save?

On a long stable prefix, up to 90% of the input portion of the bill.

Does caching change answers?

No. It changes billing and latency, not the computation.

Frequently searched

By capabilityReasoningCodingVisionLong context 1M+Free tierOpen weightsLow latencyRealtime voiceImage generation
By vendorAnthropic (13)OpenAI (30)Google DeepMind (17)xAI (8)Meta (6)DeepSeek (9)Mistral AI (11)Alibaba (Qwen) (34)Z.ai (12)Moonshot AI (5)
Popular modelsClaude Fable 5.1GPT-6 AstraClaude Opus 5Claude Fable 5GPT-5.6 SolGLM 5.3Grok 4.6Kimi K3GPT-5.6 Terra
By deliveryBatch APIPrompt cachingPriority tierFrontier tierProprietary APITool callingStructured JSONWeb searchMultilingual

Ratings 0

Avg. score 0

Rate this page

Help us improve