
What is prompt caching? Put simply, the provider stores the fixed front part of your prompt - system instructions, tool definitions, a knowledge base - and on later calls bills that part as a cache read, typically only 5-12% of the normal input price. Repeat the same context often enough and the savings show up fast, provided the cache is built correctly and you avoid two traps: cache-write fees and unstable prefixes. All prices below come from models listed on the AI computing marketplace; data as of November 18, 2026.
Three prices matter. Normal input: what you pay per token when nothing is cached. Cache write: the one-time cost of storing a prefix, usually slightly above the normal input rate. Cache read: what the repeated prefix costs on every hit, and the cheapest of the three. The savings hinge on the ratio between cache-read and input price - across our listings that ratio sits between 1/8 and 1/20.
| Model | Input | Cache read | Read / input | Effective input |
|---|---|---|---|---|
| gpt-6-astra | $10 | $1.00 | 1/10 | $0.099 |
| claude-opus-5-5 | $4 | $0.20 | 1/20 | $0.198 |
| gpt-6.1-sol | $2 | $0.10 | 1/20 | $0.0396 |
| kimi-k3.1-pro | $2 | $0.20 | 1/10 | $0.55 |
| hy4-preview | $0.896 | $0.045 | 1/20 | $0.246 |
Prices are USD per million tokens. Effective input applies the deepest seller discount on the platform. The chart puts cache-read next to normal input so the gap is obvious.
The rule is simple: the front of the request must match a previous request exactly and still be inside the cache window. Keep system prompts and tool definitions stable at the front and you will hit; timestamps, random IDs and per-turn state at the front guarantee misses. Windows and minimum cache lengths differ by provider, and an expired entry is rewritten on the next call - which is why a bill that is cheap one hour and expensive the next usually means caching is working only half the time.
First, freeze the prefix: system prompt and tool definitions should be constant text, with changing variables pushed to the end. Second, warm the cache before scaling up: send one full request after a deploy so the prefix is stored, then batch traffic picks up the cache-read rate. Third, verify in billing: cache-read tokens in the usage payload should climb with call volume. If they stay at zero, the prefix is drifting or is below the provider's minimum length.
Trap one: cache writes are not free. Storing a prefix costs a little more than a normal input - gpt-5.6 is listed with a $5 per million cache-write rate against a $4 input rate here - so one-off jobs rarely benefit. Trap two: a single changed character busts the cache. Any edit in the prefix invalidates it, so keep frequently revised text at the tail. Trap three: discounts and caching are separate levers. Most of the saving comes from the unit-price gap (reads at 5-12% of input), not from stacking seller discounts on top. Nail the cache structure first, then shop for discounts.
Rewrite your system prompt and tool definitions as a fixed prefix and send one warm-up call before batch traffic. Check the cache-read share of your usage weekly - below half means the prefix is unstable. And before switching models, run a small sample to confirm the new model caches the same way. Cache prices and discount tiers for every model are on the model plaza; the discount rules are in the pricing page.
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa
Hailuo 2.3-fast has the lowest effective price (about $0.00185/s), Kling 3.0 offers the fullest range up to 4K, and Seedance 2.5 costs most but render