✳ the wire · analysis
Prompt cache reads cost half as much on the newest AI models: 5% of the input price on Claude Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, against 10% a generation earlier
An email when a new frontier model answers on a public API, about 15 minutes after members · free, no card
On the newest flagship models from Anthropic, OpenAI and Google, reading a prompt back from the cache costs 5% of the normal input price or less: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, GPT-6.1 Terra and Gemini 4 Argon charge 5%, and Claude Fable 5.1 2.5%. The generation before charged 10%.
GENERATION BY GENERATION
- Claude Opus 5 charged $0.50 per million cached input tokens against $5 for input (10%); Claude Opus 5.5 charges $0.20 against $4 (5%).
- Claude Sonnet 5, $0.20 against $2 (10%); Claude Sonnet 5.5, $0.10 against $2 (5%).
- Claude Fable 5, $1 against $10 (10%); Claude Fable 5.1, $0.25 against $10 (2.5%).
- GPT-6 Sol, $0.20 against $2 (10%); GPT-6.1 Sol, $0.10 against $2 (5%).
- Google's five Gemini 3.x Flash models on the list charge 10%; Gemini 4 Argon charges $0.10 against $2 (5%).
ACROSS THE LIST
19 of the 40 models charge exactly 10%. The deepest discounts are Xiaomi's: MiMo-V2.6 Pro charges 0.8% ($0.0036 against $0.435) and MiMo-V2.5 2%, ahead of DeepSeek V4 Pro at 3.3%. The smallest are xAI's: Grok 4.6 and Grok 4.7 keep 25% of the input price and Grok 4.5 15%, followed by GLM-5.3 Flash at 20% and GLM-5.3 and GLM-5.2 at 18.6%.
WHAT IT SAVES
Re-reading a 100,000-token prompt from the cache costs $0.02 on Claude Opus 5.5, against $0.40 to send it as new input. Anthropic also charges to write a prompt into the cache: $5 per million tokens for a five-minute cache on Opus 5.5, 1.25 times its input price.
HOW WE COUNTED
From ArtificialWatch's LLM price list as of November 4, 2026 (artificialwatch.com/llm-pricing/prompt-caching, also at /data/llm-prices.json): each maker's own list price for input read from the prompt cache, against its normal input price, for the model's exact catalog id. Anthropic's rates were checked on its pricing page and GPT-6.1 Sol's on OpenAI's. Cache write prices and how long a cache lives differ by maker and are not compared here.
Source: ArtificialWatch prompt caching prices (artificialwatch.com/llm-pricing/prompt-caching) ↗ · the bench index
Cite this page
Writing about this story? Copy a link or a citation.
<a href="https://artificialwatch.com/wire/prompt-cache-reads-half-price-in-newest-ai-models">Prompt cache reads cost half as much on the newest AI models: 5% of the input price on Claude Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, against 10% a generation earlier</a>ArtificialWatch, "Prompt cache reads cost half as much on the newest AI models: 5% of the input price on Claude Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, against 10% a generation earlier", https://artificialwatch.com/wire/prompt-cache-reads-half-price-in-newest-ai-models
