Anthropic's Batch API halves the price of Claude Sonnet 5.5 if you can wait up to 24 hours. Prompt caching cuts the input side by 95% on every hit, in real time. On a typical month, the two end up close to the same bill, so it's worth knowing which lever you're actually pulling before you move traffic into a queue.
Anthropic list prices for Claude Sonnet 5.5 (USD per 1M tokens, checked 2026-11-05)
| Tier | Input | Cache read | Cache write (5m / 1h) | Output |
|---|---|---|---|---|
| Standard (real time) | $2.00 | $0.10 | $2.50 / $4.00 | $10.00 |
| Batch API (async, up to 24h) | $1.00 | $0.05 | $1.25 / $2.00 | $5.00 |
Cache reads on Sonnet 5.5 are 0.05x of base input, not the 0.1x most older Claude models use, and the batch discount stacks on top. US-only inference (inference_geo) adds 1.1x to everything.
One month: 100M input tokens, 10M output tokens (cache writes left out)
| Cache-hit rate | Standard | Batch |
|---|---|---|
| 0% | $300.00 | $150.00 |
| 50% | $205.00 | $102.50 |
| 80% | $148.00 | $74.00 |
| 90% | $129.00 | $64.50 |
The number to notice: Batch at 0% cache hits ($150) costs about the same as real time at 80% hits ($148). If your prompts share a long, stable prefix, fixing caching gets you the batch discount without the 24-hour wait. Batch only wins clearly when the prefix varies per request, or when you can get both.
Where the money goes once caching works
At 90% hits on the standard tier, input is $29 of the $129 and output is $100. From that point, input price barely matters. What moves the bill:
- Output length, thinking included. Output tokens include reasoning. Lowering effort on routine calls is the next lever after caching.
- Cache writes. A 5-minute write costs 1.25x base input, a 1-hour write 2x. With 10M tokens of 5-minute writes added to the 80% month above, the standard bill goes from $148 to $173.
- Prefix stability. A timestamp or request ID at the top of the system prompt resets the hit rate to zero. Put anything that changes per request after the cached block.
A quick way to decide
- Interactive or user-facing: standard tier, and spend the effort on cache hits.
- Nightly evals, backfills, bulk classification: Batch, and keep caching on anyway, since batch cache hits are best-effort but still billed at $0.05.
- Mixed: route by latency need, not by model.
This is IteraCompute. We serve anthropic/claude-sonnet-5-5 through an OpenAI-compatible API, real time, at $1.00 input / $0.05 cached input / $1.25 cache write (5m) / $5.00 output per 1M tokens (checked 2026-11-05). That matches Anthropic's Batch per-token prices without the queue: $74 for the 80%-hit month above, against $148 on Anthropic's standard tier. If your workload can wait for Anthropic's Batch API, that's the same price direct. The same key also serves Claude Haiku 5.5 and low-cost open models like DeepSeek V4.1 Flash, so a mixed route doesn't need a second client: Iteracompute Claude Sonnet 5.5
Sources: Anthropic pricing docs (model pricing, batch processing, prompt caching, data residency sections), checked 2026-11-05.
Top comments (1)
把 timestamp/request ID 放到 prompt 末尾而不是开头这个细节太值了——我们之前就是 system prompt 开头带了动态日期，缓存命中率一直上不去。另外 90% 命中时 output 占了 $100/$129，这个提醒很实在：缓存把输入打下来之后，下一步该抠的就是 output 长度和 thinking 预算了。