Claude Haiku 5.5 vs 4.5, in plain English.

Anthropic released Claude Haiku 5.5 on 2026.10.07. Haiku is the small, fast model in the Claude line: the one you wire into high-volume jobs like classification, routing, extraction, summaries and subagents, where you pay per call thousands of times a day. As of 2026.10, the new one is much cheaper per token, reads more at once and is far better at agent work. Every number below is from Anthropic's announcement and docs.

Watch the 60-second version on Instagram.

The spec sheet

It runs on the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic lists it as its fastest current model.

The price, and the catch

A million input tokens cost a dollar on Haiku 4.5. On Haiku 5.5 they cost ten cents, as long as the prompt stays at or under 100,000 tokens. Above that, both input and output switch to a higher rate that is still half of what 4.5 charged.

Anthropic's own maths: 90% cheaper under 100K, 50% cheaper above, and since about 90% of Haiku requests were under 100K, it puts the average saving at about 75%. Cache reads drop from $0.10 to $0.01 per million tokens on short prompts, and the minimum prompt length that can be cached falls from 4,096 tokens to 512, so short system prompts can now be cached at all.

The catch is the tokenizer. Haiku 5.5 uses the newer tokenizer that arrived with Claude Opus 4.7, and the docs say the same text comes out as roughly 30% more tokens than on Haiku 4.5 (the exact figure depends on the content). Per token you pay a tenth; per piece of text you pay about 0.10 × 1.3 = 0.13 of what you did. That is still roughly 87% off on short prompts, not 90%. It also means a 1M-token window holds about 555,000 words, not the 750,000 the same window held on the old tokenizer. Recount anything you budget in tokens.

The benchmarks, in plain English

Anthropic's table compares the two models on eight benchmarks. The four that report a percentage are the easiest to read:

• OSWorld 2.1 puts the model in front of a real desktop and asks it to finish tasks by clicking and typing. Haiku 4.5 finished about 1 in 6 (15.7%). Haiku 5.5 finishes about 3 in 4 (72.4%). This is the jump that changes what a small model can be trusted with.
• Humanity's Last Exam is a set of very hard expert questions. Without tools, the right-answer rate went from about 1 in 10 (10.2%) to nearly half (45.9%); with tools, from 18.7% to 57.4%.
• Chartography asks the model to read and reason over charts: 6.4% to 46.4%.
• Terminal-Bench 4.0 asks it to finish coding tasks in a shell: 0.0% to 39.2%. Anthropic doesn't explain the zero, so I would read this one as "4.5 couldn't do it; 5.5 can, some of the time" and not lean on the ratio.

The other two headline numbers need a warning. GDPval-AA (735 → 1620) and AA-Briefcase (614 → 1578) measure real professional and knowledge-work tasks, but they are Elo ratings: scores that rank models by how often one beats another in head-to-head comparisons, as in chess. An Elo of 1620 against 735 means 5.5 wins those match-ups far more often. It doesn't mean it is "2.2× as good", and the numbers aren't a percentage of anything.

Two more caveats. Haiku 5.5 has five effort levels, from low to max, and benchmark tables usually show a model at its strongest setting; the default is medium, so what you get out of the box can be lower. And these are Anthropic's own published numbers. Your evals on your own prompts are the ones that count.

What breaks when you switch

Haiku 4.5 can be retired no sooner than 2026.10.15, so this is worth doing now rather than later. Changing the model ID is not the whole migration. From the "What's new" page, the changes most likely to hit an existing Haiku integration:

• Sampling parameters can error. Omit temperature, top_p and top_k.
• Manual thinking errors. budget_tokens is gone; adaptive thinking decides when to think, and output_config.effort steers how much.
• Assistant prefill errors. End messages with a user turn.
• The first block may not be the answer. Responses can begin with thinking blocks even if you never asked for thinking. Pick blocks by type.
• Thinking counts toward max_tokens. A small limit can stop after the thinking block, before any text.
• Refusals have a stop reason. Handle stop_reason: "refusal".

A minimal call that respects all of that:

const msg = await client.messages.create({
  model: "claude-haiku-5-5",
  max_tokens: 2048, // leave room: thinking counts toward this
  messages: [{ role: "user", content: "Classify this support ticket as urgent or not: …" }],
  output_config: { effort: "low" }, // cheapest, fastest level; the default is "medium"
});
 
const text = msg.content.find((block) => block.type === "text")?.text;

The docs suggest low for chat and simple high-volume requests, medium for most work including agentic coding, and high for knowledge work and longer agent tasks. At low, long agent prompts are more likely to skip a search or stop early, so a classifier and an agent shouldn't share a setting.

The short version

1. Haiku 5.5 is a tenth of the price per token on prompts up to 100K tokens and half above. With ~30% more tokens per text, short prompts end up roughly 87% cheaper.
2. 1M tokens of context and 128K of output, up from 200K and 64K.
3. The real jump is agentic: desktop tasks from about 1 in 6 to about 3 in 4.
4. GDPval-AA and AA-Briefcase are Elo ratings. Don't divide them.
5. Drop sampling parameters, prefill and budget_tokens; read blocks by type; leave room in max_tokens. Haiku 4.5 retires no sooner than 2026.10.15.

That date matters to me: my WhatsApp agent, built on LangGraph, runs on Haiku 4.5 today, and it will probably be the first thing I move to 5.5.

Sources: Anthropic, Introducing Claude Haiku 5.5 · Claude Docs, Claude Haiku 5.5 overview · Claude Docs, What's new in Claude Haiku 5.5 · Claude Docs, Claude Haiku 4.5 overview · Claude Docs, Effort