Anthropic released Claude Haiku 5.5 on October 7. It’s the small model next to Sonnet 5.5 and Opus 5.5, pitched for high-volume work like classification, extraction, routing, compaction, and subagent calls.

What’s different is how far capability moved and how the pricing is built. On Anthropic’s table, Haiku 4.5 scored 0.0% on Terminal-Bench 4.0 and Haiku 5.5 scores 39.2%. Input is $0.10 per million tokens, a tenth of Haiku 4.5’s $1.00. Both numbers are true, and both need a footnote before you plan capacity around them.

The spec sheet

From the model overview and the what’s-new page:

  • Context window: 1M tokens, up from 200K on Haiku 4.5.
  • Max output: 128K tokens on the synchronous Messages API (up from 64K), and up to 300K on the Batch API behind the output-300k-2026-03-24 beta header.
  • Thinking: adaptive thinking, on by default, steered by an effort parameter. Default effort is medium. This is the first Haiku with effort control.
  • Modalities: text and images in, text out.
  • Knowledge cutoff: June 2026, same as Sonnet 5.5 and Opus 5.5.
  • Model ID: claude-haiku-5-5 (anthropic.claude-haiku-5-5 on Bedrock).

Anthropic doesn’t publish parameter count, architecture, or tokens per second, only that it’s the fastest model in the lineup at standard speed.

Pricing has two tiers, and that’s the design constraint

Per million tokens, from the overview:

Prompt up to 100K Prompt over 100K Haiku 4.5
Input $0.10 $0.50 $1.00
Output $0.50 $2.50 $5.00
Cache read $0.01 $0.05 $0.10
5-minute cache write $0.125 $0.625 $1.25
1-hour cache write $0.20 $1.00 —

Batch is 50% off input and output. Anthropic says about 90% of requests to Haiku 4.5 had prompts under 100K tokens, which is why it can claim the new model costs “around 75% less to run” on average. Artificial Analysis notes the under-100K rate is the same as GPT-6 Luna’s, and that the price steps up 5x above the threshold.

The 1M window is real, but it’s priced like a different product. The 100K line is where you put your chunking boundary.

The tokenizer moves that line

Haiku 5.5 uses the newer tokenizer that came in with Claude Opus 4.7. The docs say the same text produces about 30% more tokens than it did on Haiku 4.5. Anthropic’s footnote on the 75% figure says it already accounts for this. Your own budgets don’t, until you recount.

Two quick checks, by my arithmetic, using the docs’ 30% figure as a flat multiplier:

  • The tier boundary moves. A prompt that measured about 77K tokens on Haiku 4.5 is about 100K on Haiku 5.5. Anything you sized at 80 to 100K under the old tokenizer now lands on the expensive side.
  • The savings shrink above the line. A 50K-token prompt on Haiku 4.5 (about 65K on 5.5) goes from $0.050 to about $0.0065 of input, roughly 87% less. A 150K-token prompt (about 195K on 5.5) goes from $0.150 to about $0.098, roughly 35% less, not the 50% you’d read off the rate card.

Recount max_tokens, rate-limit headroom, and any dashboard that counts tokens too.

Output tokens are where the bill actually goes

Artificial Analysis scored Haiku 5.5 at 43 on its Intelligence Index at max effort. That puts it just ahead of GLM-5.3 Flash (42), Gemini 3.8 Flash (41), and GPT-6 Luna (38), close to Kimi K3 (44), and 13 points behind Sonnet 5.5 at max (56).

The more useful number in that write-up is token use. At max effort Haiku 5.5 used about 162K output tokens per Intelligence Index task, about three times GPT-6 Luna’s ~50K at max. Going from xhigh to max added 2 points for about 1.8x the tokens. At matched scores the gap is smaller but still there: Haiku 5.5 at high scored 38 with ~55K tokens per task, against Luna’s 38 at ~50K.

Since the two models have the same list price under 100K, here’s what that means in dollars, by my arithmetic: 162K output tokens at $0.50 per million is about $0.081 per task, and 50K is about $0.025. AA says its cost-per-task numbers don’t include the tier step-up yet and will follow. Until then, the effort setting matters more for cost than the rate card does. For classification and routing, I’d start at low or medium and only move up after measuring.

One mechanical catch: thinking tokens count toward max_tokens, so a small cap can end the response after a thinking block and before any text. A label classifier running with max_tokens: 256 can break.

Benchmarks, and how far apart the sources are

From Anthropic’s system card, all at max effort, averaged over five trials unless noted:

  • SWE-bench Pro: 64.8% (Sonnet 5.5: 81.3%)
  • SWE-bench Multilingual: 83.7% (Haiku 4.5: 67.4%, Sonnet 5.5: 90.3%)
  • FrontierCode 1.1 Main: 46.4% (Sonnet 5.5: 46.2% at max, 52.1% at xhigh; GPT-6 Luna: 42.4%)
  • OSWorld 2.1 offline subset: 72.4% (Haiku 4.5: 15.7%, Sonnet 5.5: 83.9%, Luna: 48.9%)
  • Humanity’s Last Exam: 45.9% without tools, 57.4% with tools
  • Terminal-Bench 4.0: 39.2%, ±1.9 standard error, 10 trials per task (Sonnet 5.5: 70.6%)

The FrontierCode tie with Sonnet 5.5 is less than it looks: Sonnet’s best Main score, at xhigh, is about six points higher.

On Terminal-Bench 4.0, Artificial Analysis measured 33%, not 39.2%: level with GLM-5.3 Flash, ahead of Gemini 3.8 Flash (20%) and Luna (13%), but six points under the vendor number. Anthropic’s run used Claude Code in --bare mode with no internet egress and resources pre-cached into the images. Different harnesses give different numbers, and only yours matches how you’ll run it.

On AA-Omniscience, accuracy is 36% against 55% for Gemini 3.8 Flash and 44% for Luna, but the hallucination rate is lower too: 40% against 55% and 77%. It says “I don’t know” more often, which is the right trade for extraction.

Safeguards show up as failed requests

Haiku 5.5 can return stop_reason: "refusal" when a safety classifier declines a request, and the docs say server-side fallback isn’t available for this model. Your client has to handle it.

The system card shows the rate. On Terminal-Bench 4.0, Anthropic ran without a fallback model, so a flagged request ended the trial: 12 of 660 trials (1.8%), ten of them on one task, all failed. On Terminal-Bench-Science it was 16 of 700, all on two tasks.

AA also flagged that its AutomationBench-AA score (35%, against 53 to 60% for Luna, Gemini 3.8 Flash, and GLM-5.3 Flash) is probably understated because of an over-refusal issue during pre-release testing that Anthropic is fixing. The system card also says it over-refused more than any other model in Anthropic’s automated behavioral audit, even though it over-refuses less than Haiku 4.5 on single-turn benign requests.

So a pipeline sending security-adjacent or ops traffic through Haiku 5.5 needs a retry path to another model, and refusals should be their own metric. Clustered on one task type, they look like a quality problem until you split them out.

Migration breaks things

Haiku 4.5 code won’t drop in unchanged. From the what’s-new table:

  • Non-default temperature, top_p, or top_k returns a 400. Remove them.
  • Manual extended thinking with budget_tokens returns an error. Use adaptive thinking and effort.
  • Assistant-message prefill returns an error. End messages with a user turn. If you used prefill to force JSON, move to structured output or a tool schema.
  • Responses can start with thinking blocks even if you didn’t ask for thinking. Select content blocks by type, not by index.
  • Editing earlier turns invalidates thinking blocks, so keep conversations append-only if you send them back.
  • Thinking blocks only replay through the account that produced them, or a linked one.
  • Computer use on the Claude API and Google Cloud needs computer_toolset_20260801 instead of computer_20250124.

Where I’d use it

Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding, and a 39.2% against 70.6% Terminal-Bench gap backs that up. I wouldn’t hand Haiku 5.5 a long, open-ended terminal task.

It fits the high-count, narrow work around a cluster: classifying alerts, pulling fields out of logs and events, summarizing a node’s recent history before a bigger model reasons about it, compacting agent transcripts. Those prompts mostly stay under 100K tokens, need little reasoning, and at low or medium effort cost little enough to run on every event instead of a sample.

Before trusting it I’d measure three things on real traffic: output tokens per call at each effort level, the refusal rate, and how many prompts cross 100K under the new tokenizer. Those decide the real cost more than the rate card does.

In the same release Anthropic cut Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens, which it says makes Sonnet about 20% cheaper on most agentic work. Where Haiku isn’t enough, that may matter more.

Create-only. Specs from Anthropic’s October 7, 2026 Claude Haiku 5.5 launch, the model overview, the what’s-new page, the system card, and Artificial Analysis.

Sources: Anthropic · Haiku 5.5 overview · what’s-new page · system card · Artificial Analysis.

Related: Claude Sonnet 5.5 · Claude Opus 5.5 · GPT-6 Sol and Luna · Mistral Large 4.

Anpoo Sivanadi, Staff Software Engineer