What Is Claude Haiku 5.5: The Fast, Cheap Claude (and Its Price Cliff)

CodexFuse Team · 2026-11-03T10:49:02.927Z · 11 min read min read

Claude Haiku 5.5 is Anthropic's fastest, cheapest small AI model — $0.10/$0.50 per 1M tokens, a 1M-token window, and an effort dial. Pricing, the 100K cliff, and where it fits.

What Is Claude Haiku 5.5, Exactly?

The short answer: Claude Haiku 5.5 is Anthropic's fastest, cheapest, and most capable small AI model — the small sibling in the Claude 5.5 family (our Claude Sonnet 5.5 explainer covers the flagship), built for speed and cost rather than maximum reasoning power. It is the first Haiku-class model with an adjustable effort dial, so you can trade thinking depth for price on every single request.

Anthropic shipped it on October 7, 2026 — that Claude Haiku 5.5 release date also anchors its support window, since Anthropic says it will stay available until at least October 7, 2027. It is available now on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure, with the model ID claude-haiku-5-5. Its knowledge cutoff is June 2026, and its outputs carry invisible text watermarks. In Anthropic's own words, it is "the cheapest, fastest, and most capable small model we've ever released" — and on average it now costs "around 75% less to run" than its predecessor.

How Much Does Claude Haiku 5.5 Cost? (The 100K-Token Price Cliff)

The short answer: Claude Haiku 5.5 pricing has two tiers — $0.10 per million input tokens and $0.50 per million output tokens, but only for prompts of 100,000 tokens or less. Cross that line and the price jumps 5x to $0.50 input and $2.50 output. That cliff is the single most misunderstood fact about this model, so everything below builds on it.

Compared with Haiku 4.5's $1 input and $5 output, the headline rate is a 90% cut — a tenth of the predecessor's price. But two caveats shrink the real-world saving. First, Haiku 5.5 ships with a new tokenizer that turns the same text into roughly 30% more tokens than Haiku 4.5 did, so you burn more tokens per request. Factoring that in, short-prompt workloads land roughly 85–90% cheaper than on 4.5, while Anthropic's own "around 75% less to run" averages across everything.

Second, the 100K cliff is decided per request: a prompt of 99,999 tokens gets the cheap rate, one of 100,001 pays the expensive one. The Batch API still halves input and output prices ($0.05/$0.25 short, $0.25/$1.25 long). Cache pricing follows the same two tiers — 5-minute cache writes at $0.125/$0.625, 1-hour writes at $0.20/$1.00. To make it concrete: agentriot's worked example puts 1 million input tokens plus 100,000 output tokens at $0.15 on Haiku 5.5, versus $1.50 on Haiku 4.5 and $3 on Sonnet 5.5.

One housekeeping note: claude.com/pricing lists subscription plans only, not API token prices — the token rates above come from Anthropic's API model docs, verified November 3, 2026. Model prices move quarterly, so confirm at the source before budgeting.

What Does the Effort Parameter Do?

The short answer: the Claude Haiku 5.5 effort parameter (set via output_config.effort) decides how hard the model thinks on a request. There are five levels — low, medium, high, xhigh, max — with medium as the API default, and thinking adapts to the task by default rather than using a fixed budget.

1. low — the cost floor. Best for subagents, background chores, and first-pass triage where volume beats depth.
2. medium — the default. The balanced setting most production traffic should start from.
3. high — noticeably stronger reasoning, still well under frontier-model cost.
4. xhigh — near the top of what a small model can do.
5. max — benchmark-leading scores, at the highest cost per request.

That last point matters more than it looks: Anthropic's headline benchmark scores are generally max-effort numbers, not default-setting numbers. Agentriot's analysis shows Terminal-Bench at 39.2% on max effort ($2.64 per attempt) versus just 20.3% at medium ($0.68 per attempt) — and Sonnet at medium scores 28.8% at that same $0.68. GDPval-AA drops from the headline 1,620 to 1,277 at medium, and the much-quoted OSWorld 72.4% is a partial-credit figure; the strict pass rate is 37.1% (Sonnet: 48.8%). None of this makes the scores dishonest — it makes them max-effort scores, and you should budget accordingly.

Manual thinking budgets are gone: thinking: {"type": "enabled", "budget_tokens": N} now returns a 400 error. You control thinking with thinking: {"type": "adaptive"} plus the effort dial instead.

Claude Haiku 5.5 vs Haiku 4.5: What's New?

The short answer: a 90% headline price cut, a 1-million-token context window (up from 200K), doubled max output (64K → 128K), and the effort dial — but migrating old code takes work, because six familiar API patterns now return 400 errors.

The six breaking changes, each returning HTTP 400:

1. budget_tokens is rejected. The API answers: "'thinking.type.enabled' is not supported for this model. Use 'thinking.type.adaptive' and 'output_config.effort' to control thinking behavior." Switch to thinking={"type": "adaptive"} plus output_config={"effort": "medium"}.
2. Non-default sampling params are rejected. Any non-default temperature, top_p, or top_k returns 400 ("temperature is deprecated for this model").
3. Assistant prefill is rejected. The trailing-assistant-turn trick now fails; use structured outputs via output_config.format instead.
4. The old computer tool is rejected. Replace computer_20250124 with computer_toolset_20260801 on the Claude API and Google Cloud, or computer_20251124 on Bedrock.
5. Editing conversation history is rejected. History is append-only by default for accounts created on or after August 31, 2026.
6. Thinking blocks are account-bound. Thinking blocks sent from another organization are silently dropped — no error, just gone.

Three silent changes ride along: thinking is on by default, thinking tokens count toward max_tokens, and there is a new refusal stop reason. Update client libraries before you migrate — the 400s are loud, but the silent ones bite.

What Is the Claude Haiku 5.5 Context Window?

The short answer: 1 million tokens of context, with up to 128K tokens of output per request — and up to 300K output tokens on the Message Batches API with the output-300k-2026-03-24 beta header.

The jump from 200K to 1M is what makes the pricing cliff above so important: a full-context request now comfortably crosses the 100K line and pays the 5x rate. The smaller headline win is cacheability — the minimum cacheable prompt dropped from 4,096 tokens to 512, so routine repeated context (system prompts, repo scaffolding) caches far more often, at $0.01 per million cached tokens on short prompts. If your workload is "same big prompt, many questions," the combination of a 1M window and a 512-token cache floor is where Haiku 5.5 quietly earns its keep.

Where Does Haiku 5.5 Fit? Subagents and the claude-haiku-5-5 API

The short answer: Haiku 5.5 is the subagent model — the cheap, fast worker you spin up by the dozen inside coding agents, while a bigger model supervises. That is the workload its price and latency are tuned for.

In Claude Code v2.1.293 or later, the haiku alias resolves to Haiku 5.5 on the Anthropic API (it still resolves to Haiku 4.5 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and Claude Platform on AWS — check your platform before assuming). A typical subagent definition pairs the alias with the cheapest effort setting:

---
name: test-log-summarizer
model: haiku
effort: low
---
Summarize test logs and report failures.

The model is picked in this order: the invocation parameter, the frontmatter model, the CLAUDE_CODE_SUBAGENT_MODEL environment variable, then the main conversation's model. New to the agent workflow? Our Claude Code beginner guide walks through the whole setup, and the Aider beginner guide covers the same cost-first agent pattern in another tool.

One caution from Anthropic's own prompting guide: at low and medium effort, the model sometimes reports a code change as done without running a check that exercises it. Instruct your subagents to run the tests or the build before reporting back — don't let the cheapest setting skip verification. Haiku 5.5 also gains the browser-use tool (browser_toolset_20260801, 27 member tools) on the Claude API and Google Cloud — something Haiku 4.5 never supported — though it is not available on Bedrock, Foundry, or Claude Managed Agents.

Claude Haiku 5.5 vs GPT-6 Luna (and the Other Flash Models)

The short answer: Haiku 5.5 currently leads the "flash model" pack on intelligence benchmarks, but GPT-6 Luna reaches nearly the same headline price more economically past 100K tokens — pick Haiku for peak scores, Luna for cost per point.

On price, Haiku 5.5 and GPT-6 Luna share the same headline short-context rate — $0.10 input, $0.50 output, $0.01 cache reads per million tokens. The divergence is the cliff: Haiku's price rises 5x past 100K input tokens, while Luna holds its base rate until 272,000 input tokens. Tokenharbor's cost-efficiency math is blunt: at max effort Haiku scored 43 versus Luna's 38, but at $0.21 per benchmark task versus Luna's $0.07 — Haiku produced 162K output tokens to Luna's 37K. Higher score, worse economy.

It is not a two-horse race everywhere: on AutomationBench-AA, DeepSeek V4.1 Flash leads at 69% while Haiku sits at 35%, and Luna matches Haiku on SciCode (55% each). For the broader model-family picture, see our Gemini 4 Argon explainer — another flagship from the same generation these flash models orbit.

What Early Users Report

The short answer: Anthropic's launch post cites four early-access customers, and all four stories rhyme: same-or-better quality at a fraction of the latency and cost.

Asana reports 30%+ latency reduction and up to 2.5x faster inference per agent turn; HubSpot cites 92.8% on a simulated-portal CRM suite; AlphaSense measured 0.84 versus 0.76 for Haiku 4.5 across 400 queries; Box saw quality 11 points higher than Haiku 4.5 at roughly half the latency. Treat these as vendor-reported launch claims, not independent benchmarks — but the direction is consistent with the subagent story: Haiku 5.5 is built for high-volume, latency-sensitive agent work where the per-request cost used to be the blocker.

Sources

• Anthropic — Introducing Claude Haiku 5.5 (https://www.anthropic.com/claude-haiku-5-5)
• Anthropic — Claude Haiku 5.5 API model overview and pricing (https://platform.claude.com/docs/en/models/haiku-5-5/overview)
• Anthropic — Claude Haiku 5.5 System Card (https://www.anthropic.com/claude-haiku-5-5-system-card)
• AWS Machine Learning Blog — Introducing Claude Haiku 5.5 on AWS (https://aws.amazon.com/blogs/machine-learning/introducing-claude-haiku-5-5-on-aws/)
• DataCamp — Claude Haiku 5.5: Features, Benchmarks, and Pricing (https://www.datacamp.com/blog/claude-haiku-5-5)
• KDnuggets — Claude Haiku 5.5: What's New and How to Use It for Fast, Cheap AI Workflows (https://www.kdnuggets.com/claude-haiku-5-5-whats-new-and-how-to-use-it-for-fast-cheap-ai-workflows)
• Magai — Claude Haiku 5.5: 90% Cheaper, and Where It Breaks (https://magai.co/claude-haiku-5-5/)
• dev.to (persimmoq) — Migrating from Haiku 4.5 to Claude Haiku 5.5: 6 Breaking Changes (https://dev.to/persimmoq/migrating-from-haiku-45-to-claude-haiku-55-6-breaking-changes-and-when-the-price-goes-up-5x-1m65)
• vktr.com — Anthropic Cuts Small-Model Pricing 90% with Claude Haiku 5.5 (https://www.vktr.com/ai-platforms/anthropic-cuts-smallmodel-pricing-90-with-claude-haiku-55/)
• Felo AI — Claude Haiku 5.5 on Felo OpenAPI (https://felo.ai/blog/claude-haiku-5-5-on-felo-openapi/)
• Agentriot — Claude Haiku 5.5: Subagents, Pricing, Effort, Benchmarks (https://agentriot.com/news/ai-models/claude-haiku-5-5-subagents-pricing-effort-benchmarks)
• TestMu AI — Claude Haiku 5.5 (https://www.testmuai.com/blog/claude-haiku-5-5/)
• TokenHarbor — Claude Haiku 5.5 vs GPT-6 Luna and China's Flash Models (https://tokenharbor.ai/blog/claude-haiku-5-5-vs-gpt-6-luna-and-china-s-flash-models)

Frequently Asked Questions

When was Claude Haiku 5.5 released?

October 7, 2026. Anthropic made it available the same day on the Claude Platform, AWS, Google Cloud, and Azure, and it will remain available until at least October 7, 2027.

Is Claude Haiku 5.5 really 90% cheaper than Haiku 4.5?

On the headline rate, yes: $0.10/$0.50 per million tokens versus Haiku 4.5's $1/$5. In practice, two things shrink the saving — prompts over 100K tokens pay 5x ($0.50/$2.50), and the new tokenizer produces roughly 30% more tokens for the same text. Expect real-world savings closer to 75–90% on short prompts, and verify current rates in Anthropic's API docs before budgeting.

What does the effort parameter do?

It sets how hard the model thinks: low, medium, high, xhigh, or max, with medium as the API default. Thinking adapts to the task automatically; manual thinking budgets (budget_tokens) are rejected with a 400 error. Note that Anthropic's headline benchmark scores are max-effort figures — medium-effort scores run lower.

Will my Haiku 4.5 code work unchanged on 5.5?

Probably not without edits. Six patterns now return HTTP 400: budget_tokens, non-default temperature/top_p/top_k, assistant prefill, the old computer_20250124 tool, editing conversation history, and thinking blocks from another organization (silently dropped). Migrate to adaptive thinking plus the effort dial first.

What is the context window of Claude Haiku 5.5?

1 million tokens of context, with up to 128K tokens of output per request (300K on the Message Batches API via a beta header). The minimum cacheable prompt dropped to 512 tokens, which makes repeated large prompts far cheaper to reuse.