Prompt caching cuts the cost of a 50-step agent run sharply for most models. In my modeled run, Claude Opus 5.5 drops from $13.04 to $1.68 per task, and GPT-6 Astra drops from $23.12 to $3.72. The saving comes from the repeated system prompt and earlier steps being billed at the cached input rate, which is a fraction of the normal input price.
Why agents benefit most
An agent resends a growing context on every step. My 50-step agent scenario starts with an 8,000-token system prompt, adds 1,500 tokens of context per step, and produces 300 visible output tokens per step over 50 steps. Each step pays again for everything before it, so most input tokens in a run are repeats. That is the workload where cached input matters most.
The totals are large. Across the models in my data, input runs from about 2.19 million to 3.10 million tokens per task, depending on the tokenizer. Output is small by comparison, at roughly 14,700 to 20,800 tokens, before reasoning. Input is where caching pays off.
Cached input prices by model
Every popular model in my price list has a separate cached input rate. The gap between normal and cached input varies a lot by provider. Prices are per 1M tokens, as of 2026-10-11:
| Model | Input | Cached input |
|---|---|---|
| MiMo-V2.6-Pro | $0.43 | $0.004 |
| DeepSeek V4.1 Flash | $0.30 | $0.006 |
| GPT-6 Luna | $0.10 | $0.01 |
| Claude Haiku 5.5 | $0.10 | $0.01 |
| Qwen3.8 Flash | $0.15 | $0.016 |
| GLM 5.3 | $0.039 | $0.038 |
| Llama 4 Maverick | $0.19 | $0.05 |
| MiniMax M3 | $0.30 | $0.06 |
| Claude Sonnet 5.5 | $2.00 | $0.10 |
| GPT-6.1 Sol | $2.00 | $0.10 |
| Claude Opus 5.5 | $4.00 | $0.20 |
| Claude Fable 5.1 | $10.00 | $0.25 |
| GPT-6 Astra | $10.00 | $1.00 |
| Grok 4.7 | $2.00 | $0.50 |
| Qwen3.8 Max Prime | $4.00 | $0.50 |
| Kimi K3 | $0.80 | $0.55 |
Two things stand out:
- Lowest cached rates: MiMo-V2.6-Pro at $0.004 and DeepSeek V4.1 Flash at $0.006.
- Smallest discounts: GLM 5.3 ($0.039 input, $0.038 cached) and Kimi K3 ($0.80 input, $0.55 cached) get little from caching on the input side.
Cheap cached input does not always mean a cheap run. Output and reasoning tokens count too, and reasoning tokens are billed as output.
Agent-50 cost with and without caching
The table uses my modeled per-task figures. "With caching" is the modeled cost per task in the 50-step agent scenario. "Without caching" is the same run priced with no cache hits.
| Model | With caching | Without caching |
|---|---|---|
| GPT-6 Luna | 4.1¢ | $0.23 |
| MiMo-V2.6-Pro | 6.5¢ | $1.00 |
| Claude Haiku 5.5 | 6.6¢ | $0.34 |
| Claude Sonnet 5.5 | 84¢ | $6.52 |
| Claude Opus 5.5 | $1.68 | $13.04 |
| Gemini 3.1 Pro Preview | $1.95 | $5.76 |
| Qwen3.8 Max Prime | $2.57 | $10.19 |
| Claude Fable 5.1 | $3.46 | $32.60 |
| GPT-6 Astra | $3.72 | $23.12 |
| Kimi K3 | $3.48 | $4.02 |
The largest absolute savings are on the expensive models: Claude Fable 5.1 ($32.60 down to $3.46), GPT-6 Astra ($23.12 down to $3.72) and Claude Opus 5.5 ($13.04 down to $1.68). The higher the normal input price, the more a cached rate saves.
Anthropic's tokenizer ratio is 1.39 against 1.00 for OpenAI, so Claude runs start with more input tokens for the same text (3,101,175 versus 2,237,500 in my data). That raises the no-cache bill, and caching claws more of it back.
Kimi K3 is the clear outlier. Its cached rate of $0.55 is high, so caching moves the run only from $4.02 to $3.48. Models with high cached prices benefit less.
The live ranking is below.
| # | Model | Cost per task | WebDev score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  DeepSeek V4.1 FlashDeepSeek | 2.4¢$24.06 / 1,000 | 1618 | 2,235,262 / 102,435 | modeled best value | |
| 2 |  GLM 5.3 FlashZ.ai | 2.6¢$25.63 / 1,000 | 1606 | 2,239,737 / 43,814 | modeled | |
| 3 |  GPT-6 LunaOpenAI | 4.1¢$40.71 / 1,000 | 1578 | 2,237,500 / 22,000 | modeled | |
| 4 |  MiMo-V2.6-ProXiaomi | 6.5¢$64.51 / 1,000 | 1625 | 2,241,975 / 24,380 | modeled best value | |
| 5 |  Claude Haiku 5.5Anthropic | 6.6¢$66.35 / 1,000 | 1585 | 3,101,175 / 50,340 | modeled | |
| 6 | GPT-5.4 NanoOpenAI | 8.7¢$86.92 / 1,000 | 1401 | 2,237,500 / 22,000 | modeled | |
| 7 |  Qwen3.8 FlashAlibaba Qwen | 9.2¢$92.31 / 1,000 | - | 2,259,875 / 96,000 | modeled | |
| 8 |  Mistral Small 4Mistral | 13¢$135 / 1,000 | - | 2,304,625 / 147,800 | modeled | |
| 9 |  Gemini 3.1 Flash LiteGoogle | 17¢$166 / 1,000 | - | 2,194,987 / 61,865 | modeled | |
| 10 |  MiniMax M3MiniMax | 18¢$182 / 1,000 | 1482 | 2,221,837 / 24,245 | modeled | |
| 11 |  Llama 4 MaverickMeta | 19¢$186 / 1,000 | 1327 | 2,241,975 / 95,880 | modeled | |
| 12 | Gemini 3.5 Flash LiteGoogle | 24¢$242 / 1,000 | 1441 | 2,194,987 / 61,865 | modeled |
What to check before you rely on caching
Caching is one part of the cost. Keep three points in mind:
- Compare with caching on, not list prices. The median across all 326 models in my 50-step agent run is 48¢. GPT-6 Luna at 4.1¢ is far below it, while Claude Opus 5.5 at $1.68 is well above it, even with caching.
- Watch output and reasoning. Qwen3.8 Flash has a low cached rate but 80,850 reasoning tokens per task. Those bill as output and can offset input savings.
- Check the basis. All 50-step agent figures here are modeled from the scenario above. The 12 measured models were run once each, at default settings, through OpenRouter, with output capped at 16,000 tokens. A single run is a guide, not a guarantee.
To estimate your own workload, use the cost calculator and set your own step count, system prompt length and output size. The methodology page explains how I build scenarios and token counts. For per-model detail, see Claude Opus 5.5 or Claude Sonnet 5.5. The scenario itself is on the 50-step agent task page.
FAQ
Does prompt caching always make agent runs cheaper?
For every model in the price table above, the cached input price is below the normal input price. Every model in the cost table also runs cheaper with caching. The size of the saving varies widely, with the biggest gaps on expensive models like Claude Fable 5.1 and GPT-6 Astra.
Which model is cheapest per agent task with caching?
GPT-6 Luna is the lowest in my 50-step agent run at 4.1¢ per task, or $40.71 per 1,000 tasks. The median across all 326 models is 48¢.
Why does Kimi K3 save so little from caching?
Its cached input rate is $0.55 per 1M tokens, close to its $0.80 normal input price. Its run costs $3.48 with caching and $4.02 without.