Claude Haiku 5.5: Price, Specs and Best Uses

Claude Haiku 5.5 is Anthropic's small, fast model. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 above that. It has a 1M-token window, 128K max output, and takes text and images. It is closed-weight and launched on 7 October 2026. Checked 31 October 2026.

Specs and prices

Rows come from Anthropic's Haiku 5.5 model page and pricing page, 31 October 2026. Haiku 5.5 is the only current Claude model priced by prompt length. A request's prompt length counts all input tokens, including cache reads and writes, and each request is priced on its own. So a request over 100,000 tokens pays the higher rate even when part of it is a cache hit. US-only inference adds 1.1x, and the multiplier applies to the higher long-prompt prices too. The minimum cacheable prompt is 512 tokens, down from 4,096 on Haiku 4.5.

Anthropic's launch post says Haiku 5.5 costs around 75 percent less to run than Haiku 4.5, which listed at $1 and $5. It also says prompts up to 100,000 tokens make up about 90 percent of requests to the previous Haiku, which is why that tier is the headline price.

What Haiku 5.5 is for

Anthropic's description is "for high-volume, latency-sensitive tasks such as classification, extraction, and routing." The launch post adds summaries, compactions, database queries, live customer support and browser use, and says Haiku "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work."

The post is also direct about limits. It says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0," and its own table shows the gap. Haiku is not a cheaper Opus. It is a different tool for work that is narrow, repeated and cheap enough to run thousands of times.

Think of three jobs where that fits:

• Triage and routing. Read an incoming ticket or log line and pick a category, a team or a next step.
• Extraction. Pull fields from documents, emails or web pages into JSON.
• Subagent lookups. In an agent run, a larger model plans and hands search, summarising or file-reading to Haiku.

Published benchmarks (vendor-reported)

Anthropic's Haiku 5.5 launch post has this table. Every figure is vendor-reported by Anthropic. The comparison models are Haiku 4.5, GPT-6 Luna and Sonnet 5.5.

The table makes the point in numbers. Haiku 5.5 is far ahead of Haiku 4.5 and ahead of GPT-6 Luna on each row where Luna is shown, and well behind Sonnet 5.5 on every row. The biggest gap is Terminal-Bench 4.0, 39.2 against 70.6. The effort level behind each cell is not stated in the text we read.

Anthropic also charts Haiku 5.5 at each effort level, from the post's data. The post's effort charts cover three benchmarks, OSWorld, GDPval-AA and Humanity's Last Exam. The scores are only in the chart graphics, so we quote none.

The customer quotes in the post are vendor-selected. They include a CRM test scoring 92.8 percent over three runs, a document-QA test of 0.84 against 0.76 for Haiku 4.5, and a coding-agent team reporting that a Haiku 5.5 sidekick keeps a top-tier FrontierCode score while cutting cost and latency. They are individual results, and we cannot generalise from them.

Behaviour changes from Haiku 4.5

The Haiku 5.5 "what's new" page lists these changes.

Two of these hit cost directly. The tokenizer change means the same prompt counts about 30 percent higher than on Haiku 4.5, and earlier thinking blocks stay in context and count as input in multi-turn conversations. Anthropic's "75 percent cheaper" is an average across its test workloads, and your savings will be smaller if your prompts are long and chatty. Safety classifiers can also decline a request, and the docs say you handle stop_reason: "refusal" in your client, since server-side fallback is not available on Haiku 5.5.

Where it is served

Anthropic lists the Claude API, Amazon Bedrock, Claude Models on Google Cloud (both variants), Microsoft Foundry and Claude Platform on AWS. Of the hosts in our directory, AWS Bedrock applies. Claude Code's haiku alias resolves to Haiku 5.5 on the Anthropic API and to Haiku 4.5 on Claude Platform on AWS, Bedrock, Google Cloud's agent platform and Foundry, per its model docs. For other options, see AI hosting.

Which coding agents can use it

Claude Code offers it as the haiku alias for simple tasks. GitHub Copilot lists Haiku 4.5 and 5.5. Cursor has a model page for Haiku 5.5. OpenCode and Cline can reach it through an Anthropic key. The Haiku launch post also names Devin's CLI, where Haiku 5.5 can act as a "sidekick" with Opus 5.5 as lead, per a customer quote. Cognition's docs were not checked. See Devin for the agent.

Cost in practice

Our arithmetic from the list prices, not a measurement. A classification call with a 2,000-token prompt and a 100-token answer costs $0.0002 for input plus $0.00005 for output, so $0.00025. A million of those costs about $250. On the Batch API it halves to about $125. The same million calls on Sonnet 5.5 would cost about $5,000 at $2 and $10, which is twenty times as much. Add reasoning tokens and the gap narrows, since adaptive thinking can add output, so set effort low for simple labels.

Long prompts change the sum. A 150,000-token document read once costs $0.075 on input at the long-prompt rate of $0.50 per million, against $0.015 if it were priced at $0.10, so the long-prompt tier is five times the short one. If the same document is asked about many times, caching helps, but each request over 100,000 tokens still pays the higher rate, including its cache reads at $0.05. For repeated questions on one large document, it can be cheaper to chunk it and keep each request under 100,000 tokens.

The subagent pattern

Anthropic's strongest pitch for Haiku 5.5 is as a worker under a larger model. The pattern is simple. A planner model, Opus 5.5 or Sonnet 5.5, breaks a job into steps. Narrow steps such as searching a repository, reading a long log or summarising a file go to Haiku. The planner reads the short results and carries on. Because the planner's context stays small, you pay Opus or Sonnet rates on a few thousand tokens instead of on every file the job touches.

Claude Code supports this through subagents that can name their own model, and the haiku alias exists for exactly that. We wrote about the mechanics on how to write a subagent. Two cautions apply. A small model can miss detail in a long file, so ask it to quote the lines it relies on. And the advisor tool rejects Haiku 5.5 as an advisor on Sonnet 5.5, so do not assume every pairing is allowed.

Haiku price history

The pricing page keeps older Haiku models listed, so you can see how far the price has moved.

Haiku 4.5 had a 200K context window and 64K max output, per the Haiku 5.5 notes. The new model has five times the window and twice the output, at a tenth of the short-prompt price. The trade, besides the long-prompt tier, is the larger token count from the newer tokenizer.

Writing prompts for a small model

Anthropic ships a prompting page for Haiku 5.5, and a few habits follow from the docs we read. Keep the instruction short and specific. State the output format, and if you need JSON, use strict tool use, since forced tool choice is not the route on the newer models. Put examples in the prompt for classification, because a label set with three examples is cheaper than a larger model. Cache the static instruction block, which is easy because the minimum is now 512 tokens. And log refusals separately from wrong answers, since a refusal is a stop reason and not a bad output.

Local hardware needs

None. Haiku 5.5 is not downloadable. For small open-weights models that run on a laptop, see open source LLMs, Qwen and Gemma.

Who should pick it

Our take

Haiku 5.5 has the most interesting price of the four Claude models. Ten cents in and fifty cents out, with cache reads at a cent, puts it in the same band as GPT-6 Luna, and Anthropic's own table has it ahead of Luna on each shared row. The 1M context is generous, though the pricing steps up past 100,000 tokens, so it is not a cheap long-document model.

Our advice is to use it where the task is small and the volume is large, to leave effort at low or medium unless a test shows a gain, and to measure token counts after the tokenizer change before quoting savings to your finance team. We have not tested the model. This view comes from Anthropic's docs and launch post. For the whole family see Claude models, and for the OpenAI counterpart see GPT models.

Checklist

• Recount tokens on real prompts, since counts rise about 30 percent against Haiku 4.5.
• Keep prompts under 100,000 tokens where you can.
• Select response blocks by type, not by position.
• Omit sampling parameters and assistant prefill.
• Handle stop_reason: "refusal" in the client.
• Set effort to low for simple labels and check quality.
• Use the Batch API for offline jobs.
• Recheck pricing before committing; this page is dated 31 October 2026.

Questions we hear

Is Haiku 5.5 good enough for coding? For subagent steps and narrow tasks, per Anthropic. For complex agentic coding, Anthropic points to Sonnet 5.5 and Opus 5.5.

Does Haiku 5.5 cost less than Sonnet? Yes, at about one twentieth of the Sonnet price on short prompts.

What replaces Haiku 4.5? Haiku 5.5 is the successor, and Haiku 4.5 stays listed at $1 and $5.