Claude Haiku 5.5: Anthropic’s Cheapest Fast Model — and the 100K-Token Price Cliff
Anthropic released Claude Haiku 5.5 on October 7, 2026. It is the third Claude 5.5 model in a month. It arrived ahead of the company’s planned IPO. It is the cheapest fast model Anthropic has shipped. It costs roughly 75% less than Haiku 4.5. But anyone comparing Claude Haiku 5.5 pricing needs to understand one thing first. There is a 100K-token price cliff. Cross it and your rate jumps fivefold on the whole request.
Key Takeaways
- Claude Haiku 5.5 pricing starts at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens.
- Prompts over 100K tokens are repriced to $0.50 input and $2.50 output per million tokens. The higher rate applies to the entire request, not just the tokens above the threshold.
- The model ships with a 1M token context window, 128K max output, and the first controllable reasoning effort setting in a Haiku model.
- It is available on Claude.ai, Claude Code, the API, AWS, GCP, and Azure.
- Developers with prompts near the boundary should monitor prompt sizes and compress context to stay under the cliff.
Claude Haiku 5.5 Pricing
| Prompt size | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Up to 100K tokens | $0.10 | $0.50 |
| Over 100K tokens | $0.50 | $2.50 |
Here is the gotcha. Crossing 100K prompt tokens reprices the entire request, not just the tokens above the threshold. A 101K-token prompt costs five times the input rate on all 101K tokens. If your prompts hover near the boundary, that cliff can quintuple a bill without warning.
What the 100K-Token Price Cliff Means in Practice
Most API pricing is simple. You pay per token. More tokens means a bigger bill in a straight line. The Haiku 5.5 cliff breaks that pattern. It creates two separate price tiers with a hard boundary between them.
Think of it as two different products sharing one name. Under 100K tokens, you get one of the cheapest capable models on the market. Over 100K tokens, you pay rates that look more like a mid-tier model. The switch flips on the whole prompt at once.
A quick worked example makes this concrete. Take a 100K-token prompt. At the lower tier, input costs 0.1 million times $0.10. That is $0.01. Now add just 1,000 more tokens. The prompt is 101K. The whole thing reprices at $0.50 per million. That is 0.101 million times $0.50, or about $0.05. One thousand extra tokens multiplied the input cost roughly five times.
Output tokens follow the same cliff. Under the boundary, output is $0.50 per million. Over it, output is $2.50 per million. Long generations on long prompts get expensive fast. This matters most for agents and summarizers that both read a lot and write a lot.
The practical lesson is simple. Treat 100K as a hard ceiling in your design, not a soft guideline. Build your prompts, your chunking, and your budgets around staying under it.
What Else Is New
- 1M token context window, with 128K max output. You can feed the model very long documents. But remember that the price cliff sits at 100K, well below the context limit.
- Effort setting. This is the first Haiku model with controllable reasoning effort. Turn it down for simple tasks where speed matters. Turn it up for trickier steps where a little extra care pays off.
- Wide availability. Haiku 5.5 is available on Claude.ai, Claude Code, the API, AWS, GCP, and Azure.
- Related change. Sonnet 5.5 cache reads were halved in price. This pairs well with a Haiku-first architecture where Haiku handles volume and Sonnet handles the hard parts.
How to Stay Under the Cliff
- Monitor prompt sizes in your logging. Alert at 90K, not 100K. System prompts, tool definitions, and conversation history all count toward the total.
- Compress context aggressively. Use summaries, retrieval, and truncation before the call. Instead of pasting a whole document into the prompt, retrieve only the passages that matter.
- Split long jobs into sub-100K chunks. Document processing is the classic case. Process chapters or sections independently, then combine the results.
- Budget with the cliff in mind. Model your worst case at the higher tier. Surprises in AI bills almost always come from the tail of the distribution, not the average.
Haiku 5.5 vs Haiku 4.5: What Actually Changes
The headline change is price. Haiku 5.5 costs roughly 75% less than Haiku 4.5. For teams running Haiku 4.5 in production, this is close to a straight upgrade at a lower price. Same fast model family. Lower bill. New effort control on top.
The effort setting is the main capability addition. Haiku 4.5 gave you one fixed behavior. Haiku 5.5 lets you tune how much reasoning goes into each call.
One thing to watch during migration: the price cliff is new. Haiku 4.5 did not have this two-tier structure. If your 4.5 prompts were already long, check where they land against the 100K boundary before you switch.
This launch also fits a broader industry trend. Every major lab is now competing on inference cost as hard as on capability. Our analysis of OpenAI and Anthropic’s cheaper 2026 models tracks how aggressively prices have been falling across the board.
Who Should Use Haiku 5.5?
- Agent and bot builders. Fast, cheap, and now with effort control for simple reasoning steps. Haiku 5.5 is built exactly for this pattern.
- High-volume classification and routing. At $0.10 per million input tokens, you can classify enormous volumes for very little money.
- Anyone on Haiku 4.5. This is a straight upgrade at a lower price. Migrate when convenient. Just audit your prompt sizes against the cliff first.
- Cost-sensitive startups. If you are watching every dollar of API spend, Haiku 5.5 stretches a budget further than any previous Anthropic model.
Who Should Think Twice
Haiku 5.5 is not for everyone. If your prompts routinely exceed 100K tokens, the higher tier erases much of the price advantage.
Teams doing deep reasoning, complex coding, or long-form generation should also look at larger models. Haiku is the fast, cheap tier. It is not meant to replace Sonnet or Opus on hard tasks. For background on how Anthropic positions its lineup, see our Anthropic overview.
Frequently Asked Questions
How much does Claude Haiku 5.5 cost?
$0.10/$0.50 per million input/output tokens for prompts up to 100K tokens. Above that, $0.50/$2.50 per million.
What is the 100K-token price cliff?
Prompts over 100K tokens are billed at five times the input rate. The higher rate applies to the whole request, not just the tokens above the threshold.
When was Claude Haiku 5.5 released?
October 7, 2026.
Does the cliff apply to output tokens too?
Yes. Output jumps from $0.50 to $2.50 per million tokens when the prompt crosses 100K.
Should I switch from Haiku 4.5?
In most cases, yes. It is roughly 75% cheaper with an added effort control. Audit your prompt sizes against the 100K cliff first.
Where can I use Haiku 5.5?
On Claude.ai, in Claude Code, through the API, and on AWS, GCP, and Azure.
