{"schema_version":"1.0","slug":"llm-cache-pricing-2026","title":"LLM Cache Pricing 2026: The Real Cost of Cached Input","description":"LLM cache pricing 2026 normalized across six providers: effective input cost at real hit rates, hidden write fees, and which vendors publish nothing at all.","data_updated":"2026-07-28","source_post":"https://www.jsonhouse.com/posts/llm-cache-pricing-2026/","category":"ai-data-statistics","cluster":"CLUSTER_LLM","format":"D","key_facts":[{"fact":"The cached-read discount converged on exactly 90% of base input at Anthropic, OpenAI, and Google; xAI discounts 84-85% and DeepSeek 98-99.2%","source":"jsonhouse normalization of six official pricing pages, 2026-07-28","category":"data"},{"fact":"Effective input price is base x (1 - h) + cache_read x h, where h is the share of input tokens served from cache","source":"jsonhouse original methodology, 2026-07-28","category":"definition"},{"fact":"The cheapest model by effective input cost changes at an 84.7% cache hit rate: Gemini 2.5 Flash-Lite below it, DeepSeek v4-flash above it, both costing $0.0237 per 1M at the crossover","source":"jsonhouse original calculation, 2026-07-28","category":"data"},{"fact":"Three unrelated model pairs cross over within an 83.3%-84.7% band, the point where a 90% discount can no longer defend a lower sticker price against a 98% one","source":"jsonhouse original calculation, 2026-07-28","category":"trend"},{"fact":"OpenAI introduced a 1.25x cache-write charge with the GPT-5.6 family; models before it had no write fee","source":"OpenAI prompt caching guide, retrieved 2026-07-28","category":"policy"},{"fact":"Anthropic charges 1.25x base input for a 5-minute cache write and 2x for a 1-hour write, and has no storage fee","source":"Anthropic pricing page, retrieved 2026-07-28","category":"policy"},{"fact":"Google charges no cache write fee but bills storage at $1.00 per 1M cached tokens per hour, rising to $4.50 on Gemini 2.5 Pro and 3.1 Pro Preview","source":"Google Gemini API pricing, retrieved 2026-07-28","category":"policy"},{"fact":"Four of the six providers publish no cache TTL, and Mistral publishes no per-model cache price at all, stating only a headline -90% for cached input","source":"jsonhouse normalization of six official documentation sets, 2026-07-28","category":"data"},{"fact":"Caching beats not caching from the second call at a 1.25x write multiplier and the third call at 2x, so write fees are immaterial for repeated workloads","source":"jsonhouse original calculation, 2026-07-28","category":"data"},{"fact":"A 50,000-token prompt sent on 1,000 Claude Sonnet 5 calls costs $100.00 uncached and $10.12 with the 5-minute cache including its 1.25x write, an 89.9% saving","source":"jsonhouse original calculation, 2026-07-28","category":"case_study"},{"fact":"Anthropic's minimum cacheable prompt length ranges from 512 tokens on Opus 5 and Fable 5 to 4,096 on Opus 4.5, Opus 4.6, and Haiku 4.5, and requests below the threshold are processed without caching and without an error","source":"Anthropic prompt caching documentation, retrieved 2026-07-28","category":"policy"},{"fact":"OpenAI's documentation does not state whether a cache write is billed on a prompt that is never reused; it says an eligible prefix 'may' be cached, and offers prompt_cache_options.mode set to explicit as an opt-out","source":"OpenAI prompt caching guide, retrieved 2026-07-28","category":"policy"}],"faq_summary":[{"q":"How much does prompt caching actually save in 2026?","a":"The discount is 90% at Anthropic, OpenAI, and Google, 84-85% at xAI, and 98-99% at DeepSeek. It applies only to the cached share of a prompt, so the realized saving is roughly hit rate times discount — a 50% hit rate against a 90% discount is a 45% cut."},{"q":"Do you pay to write to the cache?","a":"Anthropic always charges: 1.25x base input for a 5-minute write, 2x for an hour. OpenAI added the same 1.25x charge with GPT-5.6 and charged nothing before it. Google charges storage per hour instead. xAI, DeepSeek, and Mistral publish no write price."},{"q":"Which LLM API is cheapest for a cache-heavy agent?","a":"It depends on hit rate. Below roughly 83% Gemini 2.5 Flash-Lite is cheapest per effective input token; above it DeepSeek v4-flash leads, reaching $0.0097 per 1M at a 95% hit rate."},{"q":"Is prompt caching automatic or do I have to enable it?","a":"OpenAI, Google, DeepSeek, and xAI cache automatically; Anthropic requires an explicit cache_control field. Because GPT-5.6 caches automatically above 1,024 tokens and bills writes at 1.25x, a prompt you never reuse may bill above base rate unless you opt out."},{"q":"Why can't I compare cache prices straight from provider pricing pages?","a":"No two publish the same fields. Anthropic gives multipliers, DeepSeek an absolute cache-hit price, xAI one cached-input column, Google a per-hour storage rate, and Mistral only a headline -90%. Four of six publish no TTL."}],"primary_sources":[{"title":"Claude API pricing","url":"https://platform.claude.com/docs/en/docs/about-claude/pricing","publisher":"Anthropic"},{"title":"Claude prompt caching","url":"https://platform.claude.com/docs/en/docs/build-with-claude/prompt-caching","publisher":"Anthropic"},{"title":"OpenAI API pricing","url":"https://developers.openai.com/api/docs/pricing","publisher":"OpenAI"},{"title":"OpenAI prompt caching guide","url":"https://developers.openai.com/api/docs/guides/prompt-caching","publisher":"OpenAI"},{"title":"Gemini API pricing","url":"https://ai.google.dev/gemini-api/docs/pricing","publisher":"Google"},{"title":"Gemini API context caching","url":"https://ai.google.dev/gemini-api/docs/caching","publisher":"Google"},{"title":"xAI models and pricing","url":"https://docs.x.ai/docs/models","publisher":"xAI"},{"title":"xAI prompt caching usage and pricing","url":"https://docs.x.ai/developers/advanced-api-usage/prompt-caching/usage-and-pricing","publisher":"xAI"},{"title":"DeepSeek API pricing","url":"https://api-docs.deepseek.com/quick_start/pricing","publisher":"DeepSeek"},{"title":"DeepSeek context caching","url":"https://api-docs.deepseek.com/guides/kv_cache","publisher":"DeepSeek"},{"title":"Mistral API pricing","url":"https://mistral.ai/pricing/api","publisher":"Mistral AI"}],"comparison_data":{"dimensions":["provider","base_input_per_1m_usd","cache_read_per_1m_usd","discount_pct","effective_h50","effective_h80","effective_h95"],"unit":"USD per 1M input tokens, standard non-batch base tier, collected 2026-07-28. effective_hN = base x (1 - N) + cache_read x N","entries":[{"model":"Claude Opus 5","provider":"Anthropic","base_input_per_1m_usd":5.0,"cache_read_per_1m_usd":0.5,"discount_pct":90.0,"effective_h50":2.75,"effective_h80":1.4,"effective_h95":0.725},{"model":"GPT-5.6-sol","provider":"OpenAI","base_input_per_1m_usd":5.0,"cache_read_per_1m_usd":0.5,"discount_pct":90.0,"effective_h50":2.75,"effective_h80":1.4,"effective_h95":0.725},{"model":"GPT-5.6-terra","provider":"OpenAI","base_input_per_1m_usd":2.5,"cache_read_per_1m_usd":0.25,"discount_pct":90.0,"effective_h50":1.375,"effective_h80":0.7,"effective_h95":0.3625},{"model":"Claude Sonnet 5","provider":"Anthropic","base_input_per_1m_usd":2.0,"cache_read_per_1m_usd":0.2,"discount_pct":90.0,"effective_h50":1.1,"effective_h80":0.56,"effective_h95":0.29},{"model":"Grok 4.5","provider":"xAI","base_input_per_1m_usd":2.0,"cache_read_per_1m_usd":0.3,"discount_pct":85.0,"effective_h50":1.15,"effective_h80":0.64,"effective_h95":0.385},{"model":"Gemini 3.6 Flash","provider":"Google","base_input_per_1m_usd":1.5,"cache_read_per_1m_usd":0.15,"discount_pct":90.0,"effective_h50":0.825,"effective_h80":0.42,"effective_h95":0.2175},{"model":"Mistral Medium 3.5","provider":"Mistral","base_input_per_1m_usd":1.5,"cache_read_per_1m_usd":null,"discount_pct":null,"effective_h50":null,"effective_h80":null,"effective_h95":null},{"model":"Grok 4.3","provider":"xAI","base_input_per_1m_usd":1.25,"cache_read_per_1m_usd":0.2,"discount_pct":84.0,"effective_h50":0.725,"effective_h80":0.41,"effective_h95":0.2525},{"model":"Claude Haiku 4.5","provider":"Anthropic","base_input_per_1m_usd":1.0,"cache_read_per_1m_usd":0.1,"discount_pct":90.0,"effective_h50":0.55,"effective_h80":0.28,"effective_h95":0.145},{"model":"GPT-5.6-luna","provider":"OpenAI","base_input_per_1m_usd":1.0,"cache_read_per_1m_usd":0.1,"discount_pct":90.0,"effective_h50":0.55,"effective_h80":0.28,"effective_h95":0.145},{"model":"DeepSeek v4-pro","provider":"DeepSeek","base_input_per_1m_usd":0.435,"cache_read_per_1m_usd":0.003625,"discount_pct":99.2,"effective_h50":0.2193,"effective_h80":0.0899,"effective_h95":0.0252},{"model":"Gemini 2.5 Flash","provider":"Google","base_input_per_1m_usd":0.3,"cache_read_per_1m_usd":0.03,"discount_pct":90.0,"effective_h50":0.165,"effective_h80":0.084,"effective_h95":0.0435},{"model":"DeepSeek v4-flash","provider":"DeepSeek","base_input_per_1m_usd":0.14,"cache_read_per_1m_usd":0.0028,"discount_pct":98.0,"effective_h50":0.0714,"effective_h80":0.0302,"effective_h95":0.0097},{"model":"Gemini 2.5 Flash-Lite","provider":"Google","base_input_per_1m_usd":0.1,"cache_read_per_1m_usd":0.01,"discount_pct":90.0,"effective_h50":0.055,"effective_h80":0.028,"effective_h95":0.0145}]},"numerical_data":{"metrics":[{"name":"crossover_hit_rate_gemini_flash_lite_vs_deepseek_v4_flash","value":84.7,"unit":"percent","note":"Both cost $0.0237 per 1M effective input tokens at this hit rate"},{"name":"crossover_hit_rate_gemini_flash_vs_deepseek_v4_pro","value":83.7,"unit":"percent","note":"Both cost $0.0741 per 1M effective input tokens at this hit rate"},{"name":"crossover_hit_rate_grok_43_vs_gemini_36_flash","value":83.3,"unit":"percent","note":"Both cost $0.3750 per 1M effective input tokens at this hit rate"},{"name":"break_even_calls_write_multiplier_1_25x","value":2,"unit":"calls","note":"Anthropic 5-minute cache and OpenAI GPT-5.6 and later"},{"name":"break_even_calls_write_multiplier_2x","value":3,"unit":"calls","note":"Anthropic 1-hour cache"},{"name":"providers_publishing_no_cache_ttl","value":4,"unit":"of 6","note":"Google, xAI, DeepSeek, and Mistral"}]}}