Claude Haiku 5.5 Guide: Tiered API Pricing, Differences from Haiku 4.5, and Migration

What Is Claude Haiku 5.5: The New Haiku, Released October 7

Anthropic's model page states that Claude Haiku 5.5 was released on October 7, 2026, "for high-volume, latency-sensitive tasks such as classification, extraction, and routing," and it is the fastest model in the current lineup. With it, all four current models are now the new generation: Fable 5.1, Opus 5.5 and Sonnet 5.5, and now Haiku 5.5.

Current Lineup and Pricing

Sources: Anthropic's official models overview, Haiku 5.5 model page, and pricing page (checked 2026-10-28), USD. Haiku 4.5 is legacy (still available) at $1/$5.

Pricing: Two Tiers Split at 100,000 Tokens

Haiku 5.5 is the only model in the current lineup priced by prompt length. Anthropic's pricing page says Claude 4.6 and later models (except Haiku 5.5) include the full 1M context at standard pricing; on Haiku 5.5, a request whose prompt is over 100,000 tokens pays higher prices.

Three rules decide which tier applies, all from the Long context pricing section of the official pricing page:

1. Prompt length counts all input tokens, including cache reads and cache writes
2. Each request is priced on its own: a request over the threshold pays the higher prices on everything, including output and cache hits, even when much of the prompt is a cache hit
3. Earlier requests are unaffected: requests already billed keep the prices they were billed at

In other words, the threshold does not mean "only the tokens above 100K cost more." The entire request moves to the higher tier, and every upper-tier price is 5 times the lower-tier price.

Key Specs: How It Differs from Haiku 4.5

Other official specs: text and image input, text output; reliable knowledge cutoff June 2026; available on the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Claude Platform on AWS, Google Cloud (Google managed and Anthropic managed), and Microsoft Foundry. claude-haiku-5-5 is a fixed model ID with no date suffix and no separate alias.

Cost Math (Arithmetic, Not Benchmarks)

This section deliberately builds in the two things readers most often miss: the new tokenizer raises token counts, and a prompt over 100,000 tokens moves the whole request to the higher tier.

Assumptions: all token volumes below are as measured on Haiku 4.5; on Haiku 5.5, input and output are both multiplied by 1.3 (Anthropic says "approximately 30 percent," and the exact increase depends on content). Cache writes and thinking tokens are excluded; standard (non-batch) billing unless noted.

Scenario A: Short Prompts (Classification, Routing)

25,000 requests per month, each measured on Haiku 4.5 at 2,000 input and 400 output tokens, for 50 million input and 10 million output tokens per month. On Haiku 5.5 each request is about 2,600 input tokens, far below the 100K threshold.

For short prompts, Haiku 5.5 actually costs about 13% of Haiku 4.5, not 10%.

Scenario B: Long Prompts (Long-Document Q&A)

1,000 requests per month, each measured on Haiku 4.5 at 80,000 input and 2,000 output tokens, for 80 million input and 2 million output tokens per month. On Haiku 5.5 each request is about 104,000 input tokens, which crosses the 100K threshold, so the whole request is billed at the higher tier.

The wrong estimate of $9 is off from the actual $58.50 by a factor of 6.5. The actual cost is still only 65% of Haiku 4.5 and far below Sonnet 5.5, but budgeting with the wrong method produces a very large error.

If 97,500 of each request's 104,000 tokens (a shared document measured at 75,000 on Haiku 4.5) are cache hits, the best case (every request hits the cache) works out as follows: cache reads cost 97.5M × $0.05 = $4.875, uncached input costs 6.5M × $0.50 = $3.25, plus $6.50 output, for about $14.63 per month. Cache reads are billed at the upper-tier $0.05, not $0.01.

But $14.63 is the best case, where every request hits the cache, and you may not get it in practice. For requests over 100K tokens, a 5-minute cache write costs $0.625 / MTok. At 1,000 requests a month, the average is only about 1.4 requests per hour (assuming 30 days), so requests are often more than 5 minutes apart and the cache expires and has to be written again. If every request has to rewrite the cache: cache writes cost 97.5M × $0.625 = $60.94, plus $3.25 uncached input and $6.50 output, for about $70.69 per month—more than the $58.50 without caching. For low-frequency workloads, estimate your hit rate before deciding to turn caching on.

⚠️ A handy conversion: 100,000 ÷ 1.3 ≈ 76,900. A prompt that measures about 77K tokens or more on Haiku 4.5 can cross the 100K threshold on Haiku 5.5. Three more things push token counts up: multi-turn conversations keep all earlier thinking blocks and count them as input; large images cost more visual tokens (Anthropic's example: a 2,000 × 1,500 pixel image costs about 2.5 times as many as on Haiku 4.5); and adaptive thinking is on by default, with thinking tokens counting toward max_tokens. Anthropic recommends recounting with token counting and model set to claude-haiku-5-5 rather than reusing Haiku 4.5 counts.

Migrating from Haiku 4.5: What Errors or Changes Behavior

Anthropic's migration guide gives Haiku 4.5 users a 10-item checklist; below it is regrouped by whether a change breaks code outright. The first step is the model ID: change claude-haiku-4-5-20251001 (or claude-haiku-4-5) to claude-haiku-5-5. After that, these five changes make existing code fail:

1. Manual budget_tokens returns 400: change thinking: {"type": "enabled", "budget_tokens": N} to {"type": "adaptive"} (or leave it unset) and control thinking depth with effort; where you ran without thinking or with a small budget, choose a lower effort level
2. Assistant prefill returns 400: ending messages with an assistant turn for the model to continue is rejected on Haiku 5.5 even with thinking off; end with a user turn, and use structured outputs to control output format
3. Sampling parameters: temperature, top_p, and top_k can return an error; Anthropic says to omit them and guide behavior with prompting
4. The older computer use tool returns 400: replace computer_20250124 with the computer_toolset_20260801 toolset on the Claude API and Google Cloud (Google managed), or with computer_20251124 on Amazon Bedrock
5. Changing earlier turns invalidates thinking blocks: if you send thinking blocks back after changing system, tools, or earlier messages, the request returns 400, so keep conversations append-only (on accounts created before 2026-08-31 00:00 UTC, the error comes only on requests that set thinking.block_binding.prefix_mismatch_behavior)

These don't return errors, but they change how your code behaves or what you are billed:

• Recount tokens: usage and token counting results are higher, a max_tokens limit tuned for Haiku 4.5 might cut off equivalent output, and cost estimates need recomputing with the new counts and tiered prices (see the cost math above)
• The first content block can be thinking: because thinking is on by default, a response can start with thinking blocks even when the request doesn't mention thinking; code that reads the first block as the answer should select blocks by type
• A small max_tokens can return thinking and no text: thinking tokens count toward max_tokens, so a small limit can stop with stop_reason: "max_tokens" after a thinking block; raise the limit or lower effort
• Thinking text is omitted by default: thinking blocks come back with an empty thinking field and only a signature; to receive summarized thinking, set thinking: {"type": "adaptive", "display": "summarized"}
• Forced tool use works, but without thinking first: with tool_choice set to any or a named tool, the response starts with the tool call
• Safety classifiers can decline a request: handle stop_reason: "refusal"; there is no server-side fallback
• Thinking replayed across accounts is dropped: thinking blocks work only in the account that produced them (or a linked account); from another account the API drops them and the request still succeeds
• Priority Tier is not supported: organizations with a Priority Tier commitment on Haiku 4.5 need to plan capacity separately

The "after" request from the official migration guide (new thinking value plus effort):

{
  "model": "claude-haiku-5-5",
  "max_tokens": 16000,
  "thinking": { "type": "adaptive" },
  "output_config": { "effort": "medium" },
  "messages": [{ "role": "user", "content": "..." }]
}

The Haiku 4.5 "before" version used "model": "claude-haiku-4-5" with "thinking": { "type": "enabled", "budget_tokens": 8000 }. If you really need thinking off, Anthropic says you can send thinking: {"type": "disabled"} at high effort or below, but recommends effort as the better way to trade quality against speed and cost.

CloudInsight Enterprise Claude API Procurement | Local Invoices, No Foreign Credit Card Needed

Haiku 5.5's unit prices are low, but tiered pricing makes monthly bills harder to predict. What enterprises get stuck on most is often not the technology but payment, invoicing, and budget control.

Contact CloudInsight for a Claude API Enterprise Quote

We offer one-stop procurement for Claude, GPT, and Gemini, Taiwan uniform invoices, and technical support in Chinese.

How to Choose: Haiku 5.5, Haiku 4.5, or Sonnet 5.5

Haiku 4.5 is still marked Active on Anthropic's deprecations page with no retirement date announced; its retirement commitment of "not sooner than October 15, 2026" has passed, so long-running projects should schedule a migration.

FAQ: Claude Haiku 5.5

How much does Claude Haiku 5.5 cost?

For prompts up to 100,000 tokens: $0.10 per million input tokens, $0.50 per million output tokens, and $0.01 for cache reads. For prompts over 100,000 tokens: $0.50 input, $2.50 output, and $0.05 for cache reads. The Batch API is half price.

How is the 100,000-token threshold counted? Do cache hits count?

Yes. Prompt length includes all input tokens, cache reads and cache writes included, and each request is priced on its own. A request over the threshold pays the higher prices on input, output, and cache hits alike.

Is Haiku 5.5 really ten times cheaper than Haiku 4.5?

Only on unit price. The same text is about 30% more tokens on Haiku 5.5, so short prompts actually cost about 13% of Haiku 4.5; prompts over 100,000 tokens cost about 65% (see the cost math above).

Can we still use Haiku 4.5?

Yes. Anthropic lists it as legacy (still available) at $1/$5, and the deprecations page still marks it Active.

What code changes does Haiku 5.5 need?

The three most common: replace budget_tokens with adaptive thinking plus effort, remove assistant prefill and sampling parameters, and read response blocks by type.

Can thinking be turned off on Haiku 5.5?

Yes. You can send thinking: {"type": "disabled"} at high effort or below, but Anthropic recommends a lower effort level first; at lower levels the model can skip thinking entirely on simpler requests.

Conclusion: Switch Short Prompts Now, Recount Long Prompts First

Three Core Recommendations

1. Move short, high-volume work to Haiku 5.5 now: even with 30% more tokens, it costs only about 13% of Haiku 4.5
2. Recount long prompts with the new tokenizer first: requests at about 77K tokens or more on Haiku 4.5 can move entirely to the higher tier
3. Run a regression pass before going live: budget_tokens, prefill, and the computer use tool return 400, and response block order has changed

Ready to Adopt Claude Haiku 5.5?

Contact the CloudInsight sales team and we'll help plan your Haiku 5.5 and Sonnet 5.5 usage mix and handle payment and invoicing.

Join our LINE Official Account for instant model-selection consultation.

---

Further Reading

• Claude Opus 5.5 and Sonnet 5.5 Guide
• Claude API Pricing Explained
• Claude API Purchase Guide
• Claude API Tutorial
• LLM API Cost Optimization

References

• Claude Haiku 5.5 — Claude Platform Docs (checked 2026-10-28)
• What's new in Claude Haiku 5.5 — Claude Platform Docs (checked 2026-10-28)
• Claude Haiku 5.5 migration guide — Claude Platform Docs (checked 2026-10-28)
• Pricing — Claude Platform Docs (checked 2026-10-28)
• Models overview — Claude Platform Docs (checked 2026-10-28)
• Model deprecations — Claude Platform Docs