Claude Haiku 5.5 is cheap at $0.10 input and $0.50 output per 1M tokens, and it has a 1M-token context window. The price sheet also has a step at 100K prompt tokens, and it applies to the whole request. If your agent or batch job sits near that line, the step matters more than the headline price. Here's the math, using Anthropic's prices checked on 2026-10-30.
The official rates
USD per 1M tokens, global inference:
| Prompt size | Input | Cache read | 5-min cache write | Output |
|---|---|---|---|---|
| Up to 100K tokens | $0.10 | $0.01 | $0.125 | $0.50 |
| Over 100K tokens | $0.50 | $0.05 | $0.625 | $2.50 |
Once the prompt crosses 100K, everything in that request is billed at the higher row: output, cache reads and cache writes too. US-only inference costs 1.1x either row. The Batch API is half price.
One request on each side of the line
An agent turn with 90% of the prompt cached and 4K output tokens:
- 99K prompt: 9.9K x $0.10 + 89.1K x $0.01 + 4K x $0.50 = about $0.0039
- 101K prompt: 10.1K x $0.50 + 90.9K x $0.05 + 4K x $2.50 = about $0.0196
2% more prompt, about 5x the cost of the turn.
A month of traffic
100M input tokens and 10M output tokens, Anthropic list prices:
| Cache hit | All prompts under 100K | 30% of tokens in prompts over 100K |
|---|---|---|
| 0% | $15.00 | $33.00 |
| 50% | $10.50 | $23.10 |
| 80% | $7.80 | $17.16 |
| 90% | $6.90 | $15.18 |
Putting 30% of the tokens over the line multiplies the bill by 2.2x at every cache-hit rate.
Where the 100K usually comes from
- Tool and MCP definitions loaded on every turn, even in an empty session
- Compaction thresholds set as a fraction of the 1M window (35% of 1M is 350K, far past the step)
- Batch prompts that pack hundreds of records into one request
- Tokenizer differences: in one published test (The New Stack, 2026-10-23), the same prompt counted about 40% more tokens on Haiku 5.5 than on DeepSeek V4.1 Flash, so a prompt measured at 80K on another model can land over 100K
What I'd do
- Count the full billed prompt (input + cache read + cache write tokens) per request, and alert at about 90K.
- Compact or split before the step, not at the context limit.
- Split large batch prompts into chunks under 100K. The shared instructions get repeated, but with caching that repetition is cheap.
- For async work, use the Batch API, which halves both rows.
Disclosure: This is IteraCompute, an OpenAI-compatible API for many models. We list Claude Haiku 5.5 at $0.05 input, $0.005 cached input and $0.25 output per 1M tokens for prompts up to 100K, and $0.25 / $0.025 / $1.25 above that (checked 2026-10-30). That's half of the list price on both sides of the step, but the step itself is still there, so the advice above applies to us too. Current prices are on the model page: Model Page
Top comments (0)