Blog / Model watch

Claude Haiku 5.5 Pricing: $0.10 Input Under 100k, 5x Above

2026-10-18cdnping editors4 min read

Claude Haiku 5.5 launched on October 7, 2026, with the model ID claude-haiku-5-5. Its price has two tiers based on prompt length. When the prompt is 100k tokens or less, input costs $0.10 and output costs $0.50 (per million tokens). Above 100k, input costs $0.50 and output costs $2.50. Anthropic says the average running cost is about 75% lower than Haiku 4.5. Below we break down the tiers, the tokenizer change, and how to estimate your own bill.

All prices come from the official Anthropic announcement and are current as of October 2026. Your actual charges are governed by the official pricing page and your invoice.

#How is Haiku 5.5 priced?

The official footnote compares the two tiers to Haiku 4.5: requests up to 100k are about 90% cheaper, and requests above 100k are about 50% cheaper. Blended together, that gives the "about 75% on average" figure. So your real savings depend on your request-length distribution, not on a fixed 75%.

Note that above 100k the unit price is 5x. The official page only says pricing is tiered by prompt length. Whether the whole request is billed at the higher tier, and how the boundary is rounded, should be confirmed in the official pricing docs. Estimate conservatively.

#How much extra will the new tokenizer add to the bill?

Anthropic says Haiku 5.5 uses an updated tokenizer, similar to Sonnet 5.5 and Opus 5.5, so the same task uses slightly more tokens than on Haiku 4.5. No fixed ratio is given, so do not multiply an old model's token counts by the new unit price.

You can measure it yourself:

1. Take 20 to 50 representative real requests (system prompt, tool definitions, typical context). Do not use only short sentences.
2. Run the same batch through Anthropic's token-counting endpoint with both claude-haiku-4-5 and claude-haiku-5-5 (see the official docs for the exact endpoint and parameters) and record the input token counts.
3. Compute the ratio, split into two groups: "up to 100k" and "over 100k".
4. Adjust your monthly budget with that ratio: new token count × the unit price of the matching tier.

The Claude Code 2.1.293 changelog says Haiku 5.5 became the default Haiku on the API, with a 1M context window, and lists the same tiered prices (see the Claude Code changelog). Version 2.1.296 also fixed token counting failures for models like Haiku 5.5 that support only adaptive thinking when they sit behind certain gateways. If you track usage through your own gateway, upgrade and then re-check your numbers.

#How do you keep prompts under 100k?

Tiering means that adding "just a bit more context" can make the whole request cost 5x per token. Example (hypothetical numbers, for illustration only):

• 50k input + 1k output: about $0.005 + $0.0005 = $0.0055
• 120k input + 1k output: about $0.06 + $0.0025 = $0.0625

The second is about 11 times the first, while the input is only 2.4 times larger. A few things you can do:

• For retrieval tasks, cap the number of retrieved chunks, and summarize before concatenating instead of stuffing in whole documents.
• Set a compaction threshold for long conversations and summarize before you approach 100k. Haiku 5.5 itself is a good fit for this job.
• Put the stable system prompt in the cache. Cache reads cost only $0.01 per million (at 100k or less) and writes cost $0.125, so the more you reuse, the better it pays off. For how to mark cache breakpoints, see How to use Claude prompt caching to save money.
• Log the input tokens of every request and draw an alert line at 100k.

#What is it good for, and what is it not?

It belongs to the Haiku line: low price, suited to high-throughput, cost-sensitive work. Reasonable uses include summarization, context compaction, classification, and subagents in multi-agent workflows. It is the first Haiku with adjustable effort, so you can tune how much thinking it spends per task. Computer use and browser use in the Python and TypeScript SDKs are currently in beta, so validate them on a small scale before production.

For complex refactoring or long-horizon planning, compare against a larger model on your own tests before deciding, rather than looking only at unit price. To compare prices and configuration with the larger model, see Getting started with Claude Sonnet 5.5.

#How do you watch latency and availability after launch?

In the first few weeks after a model launches, latency and error rates tend to fluctuate. You can:

1. Check the cdnping Anthropic API status page, which probes the official API entry point from the Cloudflare edge every 5 minutes. Note that it measures the API entry point, not the inference latency of any particular model.
2. Record time to first token and total duration in your own application. See Breaking down AI API latency for how to split them.
3. Follow changes to the official pricing page and changelog through the changes feed, so you see price adjustments right away. See How to find out right away when an LLM API price changes for the method.
4. Use the approach in Anthropic's three new models in September and October and compare old and new models in the same time window.

#Summary: a pre-launch checklist

• Measure the old-versus-new tokenizer token ratio on real requests.
• Set a 100k alert on input tokens, and compact before 100k where needed.
• Route stable prefixes through the cache and check cache reads and writes on the invoice.
• Treat computer use and browser use as beta.
• Treat the official pricing page as the source of truth and re-check it regularly.

#Sources

• Claude Haiku 5.5, Anthropic
• Claude Code changelog (2.1.293, 2.1.296)
• Claude release notes, October 7, 2026