A single image summarising this page

Prompt Caching Pricing Compared [November 2026]

How far does the input rate drop when you reuse the same long prefix (system prompt, document set, codebase) with each vendor's cache? This page is the rate-and-terms comparison; the concept is explained in the glossary entry on prompt caching, and standard (non-cached) rates are in the LLM API pricing comparison.

●Last updated: | Every vendor page was fetched on 10 Nov 2026; quotations are verbatim and the ratio/absolute conversions are ours.

Three things to know first

  • Reads are cheap, writes are expensive. Cache reads land at 0.025x to 0.1x of the normal input rate, while writes cost 1.25x to 2x (OpenAI and Anthropic both work this way). Caching a prefix you will only send once costs you more than not caching it.
  • Discounts are quoted against the normal input rate. "90% off" means different money on a $0.10 model than on a $10 model, so the table shows both the multiplier and the absolute rate.
  • Terms differ more than prices do. Some caches are automatic (OpenAI, Moonshot, DeepSeek, Google's implicit caching), while Anthropic makes you place explicit cache breakpoints.

Cache rates by model (USD per 1M tokens)

VendorModelNormal inputCached input (hit)Read discountWrite (5 min)Write (1 hour)Minimum tokensRetention / billing unit
OpenAIGPT-6.1 Terra$1.00$0.0595% off (0.05x)$1.25$2.501,02430 min default; 12 h option in the Responses API
OpenAIGPT-6.1 Sol$2.00$0.1095% off (0.05x)$2.50$5.001,024As above
OpenAIGPT-6 Sol$2.00$0.2090% off (0.1x)$2.50$5.001,024As above
OpenAIGPT-6 Luna$0.10$0.0190% off (0.1x)$0.125$0.251,024As above
AnthropicClaude Fable 5.1 / Mythos 5.1$10.00$0.2597.5% off (0.025x)$12.50$20.00Not confirmedChoose 5 min or 1 hour at write time
AnthropicClaude Fable 5.2$8.00$0.2097.5% off (0.025x)$10.00$16.00Not confirmedChoose 5 min or 1 hour at write time (verified 2026-11-14)
AnthropicClaude Opus 5.5$4.00$0.2095% off (0.05x)$5.00$8.00Not confirmedAs above
AnthropicClaude Sonnet 5.5$2.00$0.1095% off (0.05x)$2.50$4.00Not confirmedAs above
AnthropicClaude Haiku 5.5$0.10$0.0190% off (0.1x)$0.125$0.20Not confirmedAs above
GoogleGemini 4 Argon$2.00 (to 30 Nov)$0.10 ($0.20 from 1 Dec)95% off, then 90%Not listed (implicit caching only)Not listedNot confirmedStorage $0.50/1M/hour (to 31 Dec) then $1.00
Moonshotkimi-k3.1-pro$2.00$0.2090% off (0.1x)$2.00$4.00Not confirmed5 min default; each hit refreshes the TTL
Moonshotkimi-k3.1-pro-highspeed$4.00$0.2095% off (0.05x)$4.00$8.00Not confirmedAs above
Moonshotkimi-k3$3.00$0.3090% off (0.1x)$3.00$6.00Not confirmedAs above
DeepSeekdeepseek-flash$0.15 / $0.30$0.003 / $0.00698% offNot listedNot listedNot confirmedAutomatic (off-peak / peak)
DeepSeekdeepseek-v4-pro$0.66 / $1.32$0.022 / $0.04496.7% offNot listedNot listedNot confirmedAutomatic (off-peak / peak)

OpenAI write rates are the official multipliers (1.25x for the 30-minute lifetime, 2.5x for 12 hours) applied to normal input; the docs stress that a cache write is not an additive fee but a replacement rate. Anthropic writes are 1.25x (5 min) and 2x (1 hour), reads are 0.1x (0.025x on Fable 5.1 and Mythos 5.1, 0.05x on Opus 5.5 and Sonnet 5.5). DeepSeek shows off-peak / peak.

Write versus read: when caching pays off

OpenAI states the break-even arithmetic verbatim:

"At the 0.1× read rate with 30-minute writes, writing a prefix once and fully reusing it once costs 1.35× its ordinary input cost, compared with 2× for processing it twice without caching."

The same structure holds elsewhere. For Moonshot's kimi-k3.1-pro a 5-minute write ($2.00) equals the normal input rate ($2.00) while reads fall to $0.20. Anthropic charges 1.25x or 2x on writes in exchange for reads down to 0.025x-0.1x. How many times you reuse the prefix is what decides the bill.

What we could not confirm (we do not guess)

ItemStatusNote
Anthropic minimum cacheable lengthNot confirmedNeither the pricing page nor the prompt-caching docs we fetched on 2026-11-14 state a minimum token count. We will add per-model values from the model docs once verified.
Per-model attribution of Google cache ratesNot confirmedModel names sit outside the pricing tables on ai.google.dev, so quoting a cache row alone does not establish which model it belongs to. We associate the Gemini 4 Argon row because its input/output rates ($2.00/$10.00) match that model's table.
xAI (Grok) cache rateNot confirmedThe official models page yielded only context length and input/output rates ($2.00/$6.00) verbatim; no cached-input row was retrievable.
Qwen (Alibaba) cache rateNot confirmedWe could not obtain a verbatim quote from the official pricing page for this release; it will be added in a later update.
DeepSeek write charge and minimum lengthNot confirmedThe official table we checked lists only cache-hit and cache-miss rows; no separate write or minimum-cache-length row was found (we do not claim there is no write charge).

Sources (all fetched 10 Nov 2026)

Related pages