Caching changes which model is cheapest
Most price tables sort by input price. That's the wrong number if your app resends the same system prompt, tool definitions or chat history on every call, because then most of your input is billed at the cached rate, and providers discount cached tokens by very different amounts: from 75% off to 98% off in this table.
Ranked by input price and then by the cache-heavy price, these models gain the most places:
- DeepSeek V4 Pro: 42 → 30, because cached input is 96.7% off
- DeepSeek V4.1 Flash: 16 → 12, because cached input is 98% off
- Mistral Large 3: 22 → 18, because cached input is 90% off
And these lose the most, either because their cache discount is smaller than their neighbours' or because their output price is high for their input price, and caching does nothing for output:
- OpenAI GPT-5.3 Codex: 47 → 59 (90% off cached input)
- Moonshot AI Kimi K2.7 Code: 32 → 41 (80% off cached input)
- Moonshot AI Kimi K2.6: 31 → 39 (83.2% off cached input)
If you're choosing between two models for an agent or a long-context chat, compare them on the cached price, not the headline. The model switch cost calculator does that for your own token counts.
Cache writes: who charges extra
Anthropic, Moonshot AI, and OpenAI list a separate price for writing tokens into the cache on 22 of these models. Anthropic charges 1.25x the input price for a 5-minute cache and 2x for a 1-hour cache. The “Cache write” column shows the shortest option. If the fresh 10% of your traffic is written to the cache on every call, the cache-heavy price goes up a little, for example:
- Anthropic Claude Haiku 5.5: $0.0435 → $0.046
- OpenAI GPT-6 Luna: $0.0435 → $0.046
- OpenAI GPT-5.6 Luna: $0.097 → $0.102
- OpenAI GPT-6.1 Terra: $0.393 → $0.417
On most other models caching is automatic and there is no write charge listed. Google's explicit context caching for Gemini is the exception that isn't in this table: it also bills storage per million tokens per hour, which depends on how long you keep the cache.
What this table leaves out
- Cache hit rate. These prices assume the cache hits. Real hit rates depend on the provider's minimum prefix size, cache lifetime and how stable your prompt prefix is.
- Long prompts. Models marked “long prompts cost more” charge a higher rate above a context threshold, cached tokens included.
- Off-peak pricing. DeepSeek V4.1 Flash and DeepSeek V4 Pro cost half as much outside DeepSeek's peak hours, shown as the off-peak price under the cache-heavy column. The main figure is the peak price.
- Price changes. Cached rates move without much notice: half of the price increases we have recorded since 2024 touched only cached-input prices.
The prompt caching savings calculator prices your own workload: cache lifetime and write costs for Anthropic and current OpenAI models, and a hit rate you enter for other providers. The increases are covered in do LLM API prices go down?, and every recorded change is in the LLM API price history. For headline prices without caching, see the model pricing comparison.
Frequently asked questions
Which LLM API has the cheapest cached input?
On a cache-heavy workload (85% cached input, 10% fresh input, 5% output), Nova Micro is the cheapest model here at $0.0179 per 1M tokens. The biggest discount on a cache hit is DeepSeek V4.1 Flash, where cached input costs 98% less than regular input.
How much cheaper is cached input than regular input?
Between 75% and 98% cheaper on the 75 models listed. 11 models take 95% or more off, 38 take 90% to 95% off, and 26 take 75% to 90% off. Output tokens are never discounted, so the saving on a whole bill is smaller: 44% to 76.5% on the cache-heavy mix used here.
Do I pay extra to write to the prompt cache?
With some providers. In this table, Anthropic, Moonshot AI, and OpenAI list a separate cache write price for 22 models, for example $2.50 per 1M tokens for Claude Sonnet 5.5 against $2.00 for regular input. The other models have no write price listed on their pricing pages, which usually means caching is automatic and writes cost the same as regular input. Gemini's explicit context caching also charges for storage per hour, which is not in this table.
What does the cache-heavy price mean?
It's the blended cost of 1M tokens of traffic where 85% is input read from the cache, 10% is new input and 5% is output. That is roughly what an agent or chat app looks like when it sends the same long system prompt, tools and history on every call. Your own mix will differ. The prompt caching savings calculator prices your own workload: cache lifetime and writes for Anthropic and current OpenAI models, and a hit rate you enter for other providers.
Why is a model with a low input price not the cheapest with caching?
Because providers discount cached tokens by different amounts. A model with a slightly higher input price but a 95% cache discount can end up cheaper than one with a lower input price and a smaller discount, once most of your input comes from the cache. The table is sorted by the cache-heavy price so you can see this directly.