Claude Haiku 5.5 vs Haiku 4.5: What Changes When You Upgrade

Claude Haiku 5.5 replaces Haiku 4.5 as Anthropic's small, fast model. It scores far higher on every published test and costs less per token. But it isn't a drop-in swap, because it thinks by default. It also rejects several settings Haiku 4.5 accepted and adds safety refusals Haiku 4.5 never had.

Key Takeaways

• Haiku 5.5 beats Haiku 4.5 on every test Anthropic published, often by a wide margin, and an independent index agrees.
• It writes about twice as fast once it starts, but thinking can delay the first word by several seconds.
• Per token, it costs 90% less on prompts up to 100,000 tokens. Our Haiku 5.5 cost breakdown covers what that means per task.
• Several Haiku 4.5 habits now return errors: thinking budgets, prefill, custom temperature and the old computer-use tool.
• Haiku 4.5 is still active, so test first.

Haiku 5.5 vs Haiku 4.5 at a Glance

Here's the side-by-side, from Anthropic's model overview, its migration guide and its pricing page.

Feature	Claude Haiku 5.5	Claude Haiku 4.5
Released	October 7, 2026	October 2025
Claude API model ID	claude-haiku-5-5	claude-haiku-4-5-20251001
Context window	1M tokens	200K tokens
Max output	128K tokens	64K tokens
Price per million tokens	$0.10 / $0.50 up to 100K, $0.50 / $2.50 above	$1 / $5
Thinking	Adaptive, on by default	Manual budget, off by default
Effort levels	Low, medium, high, xhigh, max	None
Tokenizer	Newer, about 30% more tokens	Older
Temperature, top_p, top_k	Omit (custom values return errors)	Supported
Assistant prefill	Returns an error	Supported with thinking off
Safety refusal stop reason	Yes, no server-side fallback	No
Status on the Claude API	Active, until at least October 7, 2027	Active, until at least October 15, 2026

Two rows are easy to miss. Haiku 5.5 reads images at a higher resolution, and the docs say a 2,000 by 1,500 pixel image costs about 2.5 times as many visual tokens as on Haiku 4.5. And it's the first Haiku model with effort levels, which changes how you tune it.

How Much Smarter Is Haiku 5.5 vs Haiku 4.5?

A lot, on paper and in independent tests.

Vendor claim: These are Anthropic's own scores from its launch page.

Benchmark	Haiku 5.5	Haiku 4.5
OSWorld 2.1, offline subset (computer use)	72.4%	15.7%
Terminal-Bench 4.0 (agentic coding)	39.2%	0.0%
GDPval-AA v2.1 (knowledge work, Elo)	1620	735
AA-Briefcase v1.1 (knowledge work, Elo)	1578	614
Humanity's Last Exam, with tools	57.4%	18.7%
Chartography, no tools (visual reasoning)	46.4%	6.4%

The biggest jumps are in computer use and agent work, where Haiku 4.5 barely registered. That fits Anthropic's pitch for Haiku 5.5 as a subagent and browser worker.

Independent data: Artificial Analysis scores Haiku 5.5 on its Intelligence Index at 29 on low effort, 34 on medium, 38 on high, 41 on xhigh and 43 on max. Its Haiku 4.5 page shows an estimated 15 for Haiku 4.5 without reasoning, with an independent run still to come. Even Haiku 5.5 at its lowest setting nearly doubles that estimate.

Hands-on testing is more mixed. A Decrypt reviewer found Haiku 5.5 very fast but saw it miss a simple logic question. Bigger scores don't mean every answer is right.

Haiku 5.5 vs Haiku 4.5 on Speed: Which Is Faster?

It depends on which kind of fast you need.

Writing speed: Artificial Analysis measures Haiku 5.5 at 168 to 237 output tokens per second, depending on effort. Haiku 4.5 runs at about 88 tokens per second on Anthropic's API. Once Haiku 5.5 starts writing, it's roughly twice as fast or more.

Time to the first word: This is where thinking costs you. Artificial Analysis lists about 9.2 seconds to the first answer token for Haiku 5.5 at low effort, because the model thinks first. Haiku 4.5, with no reasoning, shows a time to first token of 0.52 seconds.

Those two figures don't measure exactly the same thing. Still, the direction is clear. A live chat or voice app that needs a reply to start in under a second should test with thinking turned off, which Haiku 5.5 allows at low, medium and high effort.

Customer reports on Anthropic's page point to faster finished tasks overall. Asana reported task completion latency more than 30% lower than its current model, and Box said Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency. These are vendor-chosen quotes, so test your own flows.

Which Effort Level Matches Haiku 4.5?

There's no direct match, because Haiku 4.5 had no effort setting. Anthropic's prompting guide gives this map instead.

Effort	Intelligence Index (AA)	Anthropic's guidance
Low	29	Chat, short tool tasks, simple high-volume requests
Medium (default)	34	Start here for most work, including agentic coding
High	38	Knowledge work, longer agent tasks, strict instruction following
Xhigh and max	41 and 43	Only where your tests show the gain is worth it; also test Sonnet 5.5

If you ran Haiku 4.5 with no thinking, or with a small budget, Anthropic says to start at a lower effort level. Don't try to switch thinking off with a prompt. In Anthropic's testing, telling the model to answer directly didn't stop it from thinking.

The cost gap between levels is wide. Artificial Analysis puts Haiku 5.5's cheapest Intelligence Index task at about $0.02 on low effort, and its headline figure is $0.21 per task. Moving from low to medium also more than doubled output tokens per attempt in one of Anthropic's own agent tests.

Haiku 5.5 vs Haiku 4.5 on Cost

Per token, Haiku 5.5 is cheaper on every line of the price list. Per task, the gap is smaller, and a few new rules decide by how much.

Cost item	Claude Haiku 5.5	Claude Haiku 4.5
Input, per million tokens	$0.10 (prompt up to 100K)	$1.00
Output, per million tokens	$0.50 (prompt up to 100K)	$5.00
Cache read, per million tokens	$0.01	$0.10
Smallest cacheable prompt	512 tokens	4,096 tokens
Same text, token count	About 30% higher	Baseline
Large image (2,000 by 1,500 pixels)	About 2.5 times the visual tokens	Baseline

Three things shrink the saving. Thinking tokens are billed as output. The same text counts as more tokens. And any prompt over 100,000 tokens pays five times the base rate on the whole request, which is a rule Haiku 4.5 never had.

Anthropic's own estimate is about 75% lower running cost on average. In our Haiku 5.5 cost breakdown, a simple ticket-sorting job came out 83% to 87% cheaper in our own math. Long prompts, heavy thinking and big images will land lower, so measure cost per finished task, not per token.

The smaller cache minimum is a quiet win. A 1,000-token system prompt couldn't be cached on Haiku 4.5, and on Haiku 5.5 it can, at a tenth of the old cache-read price.

Which Workloads Gain Most From Haiku 5.5?

The scores above point to a pattern. Our read is that the biggest gains come where Haiku 4.5 struggled most.

• Computer and browser use: OSWorld jumps from 15.7% to 72.4%, and Haiku 5.5 adds a browser use tool that Haiku 4.5 doesn't support.
• Agent and subagent steps: Terminal-Bench goes from 0% to 39.2%, so multi-step tool work is now realistic on the small tier.
• Charts and documents: chart reading rises from 6.4% to 46.4%, and the 1M context window fits far longer files.

Simple classification and routing gain less in quality, because Haiku 4.5 often handled them well already. There, the main reason to move is cost, plus the effort setting for the harder cases.

What Breaks When You Upgrade From Haiku 4.5?

Anthropic's migration guide lists 11 changes for code moving from Haiku 4.5. These are the ones that return errors:

• Thinking budgets: sending budget_tokens returns a 400 error. Use adaptive thinking and the effort setting.
• Sampling settings: temperature must be 1 and top_p must be 0.99 if you send them at all. Any top_k value returns an error, and the Python SDK from version 1.0 drops all three.
• Assistant prefill: ending the messages with an assistant turn returns an error, even with thinking off.
• Old computer-use tool: computer_20250124 returns an error. Move to computer_toolset_20260801 on the Claude API and Google Cloud, or computer_20251124 on Amazon Bedrock.
• Edited history: if you change the system prompt, tools or earlier turns and send thinking blocks back, the request can fail. Keep conversations append-only.

Here's the thinking change in practice. Before, a Haiku 4.5 request set a budget:

{ "model": "claude-haiku-4-5", "thinking": { "type": "enabled", "budget_tokens": 8000 } }

After, the Haiku 5.5 request uses adaptive thinking and an effort level:

{ "model": "claude-haiku-5-5", "thinking": { "type": "adaptive" }, "output_config": { "effort": "medium" } }

Other changes won't throw errors but can quietly break things:

• Responses can start with thinking blocks, so pick the answer by block type, not position.
• Thinking tokens count toward max_tokens. A small limit tuned for Haiku 4.5 can end a reply before any text appears.
• Thinking text is now empty by default. Haiku 4.5 returned a summary.
• Thinking blocks from earlier turns stay in context, so long chats use more input tokens.
• Prompts count about 30% more tokens, and large images can cost far more.

For prefill, Anthropic suggests replacements by purpose. Use structured outputs or a tool with enum fields for formats, ask for a direct answer in the system prompt instead of a preamble, and move continuations into the user message.

If you use Claude Code, Anthropic's bundled API skill can do much of this for you. Running /claude-api migrate swaps the model ID, applies the breaking changes and returns a checklist to verify by hand.

How Does Haiku 5.5 Behave Differently?

Anthropic says existing Haiku 4.5 prompts "should perform well without changes." Its prompting guide then lists the cases where they don't, and most of them show up at low effort.

• Search: give the model today's date when it has a search tool. Anthropic found this grounded answers in recent results.
• JSON with tools: with thinking off and structured outputs on, the model might skip a tool call it needs. Keep thinking on for those requests, or add one line telling it to call tools first.
• Early stopping: with a long agent prompt at low effort, it sometimes stops and hands the task back. Raising effort to medium roughly halved this in Anthropic's tests.
• Unchecked code: at low and medium effort, it sometimes reports a change as done without running a check.
• Mid-task messages: it treats text inside tool results as untrusted, so a user's mid-task message placed there can be ignored. Send it as a user turn instead.
• Reasoning in replies: with thinking off or at low effort, reasoning-like text can leak into what users see.

Two more need attention. At xhigh effort in multi-turn chats, the model sometimes puts its whole answer in its thinking and returns no visible text, so check for empty replies. And safety refusals are new: a declined request comes back with a refusal stop reason and a category such as cyber, bio or general harms.

Anthropic notes that resending a refused request to Haiku 5.5 "usually returns another refusal." With no server-side fallback for this model, your code needs a plan, such as routing the request to another model or to a person.

When Should You Stay on Haiku 4.5 for Now?

For most teams, not for long. But a few cases justify waiting a few weeks.

• Priority Tier: if you have a Priority Tier commitment on Haiku 4.5, Anthropic says it isn't supported on Haiku 5.5. Plan capacity first.
• Structured outputs on Amazon Bedrock: these aren't available for Haiku 5.5 on Bedrock. You'd need prompt-described formats and your own validation.
• Sub-second first tokens: if thinking-off tests on Haiku 5.5 can't match your latency budget, keep Haiku 4.5 on that path while you test.
• Unmigrated code: pipelines that rely on prefill, fixed temperature or exact thinking budgets need changes before they switch.
• Image-heavy jobs: at 2.5 times the visual tokens for a large image, per-image savings shrink. Our math: 2.5 times the tokens at a tenth of the price still works out to about 75% less per image under 100,000 tokens.

You don't have to rush. Anthropic's deprecation table lists Haiku 4.5 as active as of October 14, 2026, with retirement not sooner than October 15, 2026. No deprecation has been announced, and Anthropic says it gives at least 60 days' notice before retiring a public model.

That date is close, though. Treat it as a reason to start testing now rather than wait for the notice.

How to Upgrade From Haiku 4.5 to Haiku 5.5 Safely

Run the new model in the shadow of the old one before you switch any traffic. Here's the order we use.

1. Apply the code changes above in a branch, and recount your prompts with the token counting endpoint using the new model ID.
2. Replay 200 to 500 real requests through Haiku 5.5 at low, medium and high effort, next to the stored Haiku 4.5 answers.
3. Track five numbers for each setting: accuracy, time to first token, total time, cost per task and refusal rate.
4. Switch one lane at a time, starting with the highest-volume, lowest-risk job, and keep Haiku 4.5 as the fallback until the numbers hold.

Most teams can do this in a week. The hardest part is usually prefill and temperature, because those choices tend to be scattered across many prompts.

Haiku 5.5 vs Haiku 4.5: Our Verdict

Upgrade, but on purpose. On quality, Haiku 5.5 isn't close to Haiku 4.5; it's in a different class, and per token it's far cheaper too.

The trade-offs are real, though. You give up sampling controls and prefill. You take on thinking and its latency, and you need a plan for refusals. None of that is a reason to stay on Haiku 4.5 for good, but each one is a reason to test before you move.

If a task needs more than Haiku 5.5 can give at high effort, look up, not back. Our Sonnet 5.5 guide covers the next tier, and our model portability lesson explains why a thin routing layer makes switches like this one cheaper.

How Van Data Team Helps Teams Upgrade Claude Models

We help teams move between model versions without surprises. That means replay tests on real traffic, side-by-side scoring, latency and cost tracking and a staged rollout with a clear way back.

Haiku 5.5 is a big step up from Haiku 4.5. Whether your prompts, code and latency budget are ready for it is something a one-week test will tell you. If you want help planning that test, our AI governance guide is a good place to start.