Table of Contents
Claude Haiku 5.5 explained: Anthropic released Claude Haiku 5.5 on October 7, 2026, the cheapest, fastest small AI model the company has ever shipped. At just $0.10 per million input tokens and $0.50 per million output tokens, it costs exactly the same as OpenAI’s GPT-6 Luna — yet in Anthropic’s own benchmark table, Claude Haiku 5.5 outscores Luna on every shared test. For US developers and startups running AI workloads thousands of times a day, that combination of price and performance is worth a close look.
In this Claude Haiku 5.5 explained guide, I break down what the model actually is, how its two-tier pricing works (including the 100K-token catch), where it beats GPT-6 Luna, where it does not, and how you can try it today.
Claude Haiku 5.5 Explained: What Is It?
Claude Haiku 5.5 is the smallest, fastest member of Anthropic’s Claude 5.5 model family. While Opus 5.5 and Sonnet 5.5 handle the hardest reasoning and agentic coding work, Haiku 5.5 is built for jobs you run at massive volume: summarizing documents, classifying tickets, compacting long conversations, answering database questions, and working as a subagent inside larger AI coding setups. This Claude Haiku 5.5 explained overview starts with the basics: the model is built for throughput, not marathon reasoning.
This release is a major step up from the previous generation. The context window jumps from 200,000 to 1 million tokens, and maximum output doubles from 64K to 128K tokens — matching the big Sonnet 5.5 and Opus 5.5 on paper. It is also the first Haiku-class model with an adjustable effort setting: a dial that lets you trade thinking time against cost and speed. Adaptive thinking is on by default, so the model decides on its own how much reasoning a task deserves, and you can override it when you want tighter control over latency or the bill.

Claude Haiku 5.5 Pricing: The 100K-Token Catch
The headline number — $0.10 per million input tokens — is real, but it comes with a condition. Anthropic prices the model on a two-tier system based on prompt length. Prompts up to 100,000 tokens get the headline rate; anything longer pays a higher row. Here is the full breakdown, per million tokens: Any Claude Haiku 5.5 explained pricing breakdown has to start with this two-tier catch.
| Prompts up to 100K tokens | Prompts over 100K tokens | |
|---|---|---|
| Input | $0.10 | $0.50 |
| Output | $0.50 | $2.50 |
| Cache read | $0.01 | $0.05 |
| Cache write (5 min) | $0.125 | $0.625 |
| Cache write (1 hour) | $0.20 | $1.00 |
Compare that with Claude Haiku 4.5 at $1 input / $5 output, and the upgrade is dramatic: list prices are 90% lower below the threshold and 50% lower above it. Anthropic says it costs around 75% less to run than Haiku 4.5 on average, noting that roughly 90% of Haiku 4.5 requests fell under the 100K mark. This Claude Haiku 5.5 explained price comparison shows the upgrade is dramatic.
Three pricing details worth knowing — these Claude Haiku 5.5 explained pricing notes cover the fine print:
- The Batch API takes 50% off input and output, which makes overnight bulk jobs — classification, backfill, document processing — extremely cheap.
- The threshold counts every input token, including cache reads and writes, and each request is priced on its own. A request that crosses 100K pays the higher row for the whole request, even the part that was a cache hit.
- The new tokenizer uses slightly more tokens per task than the old one, so it is easier to bump into the threshold than the raw character count suggests.
Anthropic also sweetened the rest of the lineup on launch day: Sonnet 5.5 cache reads were halved from $0.20 to $0.10 per million tokens, which Anthropic says makes most agentic work about 20% cheaper, and Max and Team subscribers got a new monthly API credit.
How Claude Haiku 5.5 Compares on Price
At $0.10/$0.50, Claude Haiku 5.5 matches OpenAI’s GPT-6 Luna cent for cent (Luna’s tier extends to 272K tokens, versus Haiku 5.5’s 100K cliff). If you are already building on the GPT-6 family and curious how the two models stack up, our GPT-6 Intelligent UI explained guide covers the Sol/Luna split and current US plan prices. And if you want the frontier-model picture, see our breakdown of Gemini 4 Argon, Google’s 1M-token answer to GPT-6.
Claude Haiku 5.5 Benchmarks vs GPT-6 Luna
Price is only half the story. In Anthropic’s launch-day comparison, the new Haiku beats GPT-6 Luna on every benchmark the two models share — and the gap is widest on tasks where the model has to operate a computer or a terminal. Here is Claude Haiku 5.5 explained in benchmark numbers: this Claude Haiku 5.5 explained benchmark roundup uses Anthropic’s published launch figures.
| Benchmark | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| OSWorld 2.1, offline subset (operating a desktop) | 72.4% | 48.9% |
| Terminal-Bench 4.0 (command-line tasks) | 39.2% | 16.4% |
| FrontierCode 1.1 Main (coding) | 46.4% | 42.4% |
| GDPval-AA v2.1 (office work, rating) | 1,620 | 1,437 |
The Terminal-Bench jump is the eye-catcher: Haiku 4.5 scored 0.0% on it, and Haiku 5.5 lands at 39.2%. That is exactly the kind of leap that matters for subagent coding workflows, where a small model runs shell commands, reads tool output, and summarizes results back to a bigger model. This Claude Haiku 5.5 explained benchmark highlight shows real progress.
A word of caution: all of these scores come from Anthropic’s own announcement. Anthropic does not say whether it ran GPT-6 Luna itself or used OpenAI’s published figures, so treat the comparison as the vendor’s best case rather than an independent review. Independent benchmarks will arrive in the coming weeks, and I will update this guide when they do.
It is also worth noting where it does not lead. Sonnet 5.5 still tops every row — 70.6% on Terminal-Bench 4.0, for example — and Anthropic itself points complex agentic coding work at Sonnet 5.5 and Opus 5.5. Haiku 5.5 is the fast worker, not the deep thinker.

Where Claude Haiku 5.5 Fits (and Where It Doesn’t)
The model shines wherever speed and volume matter more than maximum reasoning depth — that is Claude Haiku 5.5 explained in one sentence:
- Coding subagents: summaries, compaction, and tool-result summarization inside Claude Code, where the model gets called thousands of times a day.
- High-volume classification: support-ticket triage, content moderation, spam filtering.
- Document pipelines: summarization, extraction, and Q&A over large document sets. If your pipeline starts with speech, pair it with a strong transcription tool — see our guide to the best AI transcription software of 2026.
- Live customer support and browser-use apps, where latency directly affects the user experience.
What it is not for: long-horizon agentic coding, deep research, or tasks where a wrong answer is expensive. Keep Sonnet 5.5 or Opus 5.5 for those.
Breaking Changes If You Migrate From Haiku 4.5
Moving up is not just a model-ID swap. A thinking budget, a custom temperature, or an assistant prefill now returns a 400 error instead of working silently — those parameters are no longer accepted on Haiku 5.5. And remember the token-count cliff: because of the new tokenizer, some requests that sat comfortably under 100K on Haiku 4.5 may land above it now. Test your heaviest prompts against the threshold before assuming the headline price. Our Claude Haiku 5.5 explained migration notes flag these as the first things to test.
How to Try Claude Haiku 5.5 Today
Claude Haiku 5.5 explained for practitioners: it launched broadly on day one. You can call it on the Claude Platform with the model ID claude-haiku-5-5, use it on Claude.ai (Free, Pro, Max, Team, and Enterprise plans), or reach it through Amazon Bedrock, Google Cloud, and Microsoft’s AI platform.
Notably, this is not an API-only story. GitHub confirmed the model is generally available in GitHub Copilot (Pro, Pro+, Max, Business, and Enterprise tiers), positioned for subagents, quick edits, and terminal tasks — exactly the jobs a fast small model should own. Vercel added it to AI Gateway as well, so it slots into IDE substeps, compaction, and batchy support flows with minimal setup. A detailed practitioner walkthrough of the release, migration notes, and availability is available at DevelopersDigest’s Claude Haiku 5.5 guide, and AppStack Insider has a deep dive on the two-tier pricing if you are budgeting inference at scale.
Claude Haiku 5.5 Explained: FAQ
When was Claude Haiku 5.5 released?
October 7, 2026. Anthropic had named it as part of the 5.5 family twice in September before announcing the date. This Claude Haiku 5.5 explained FAQ entry covers the key launch facts.
How much does Claude Haiku 5.5 cost?
$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Prompts over 100K tokens cost $0.50 input / $2.50 output per million. Cache reads are $0.01 per million tokens, and the Batch API takes 50% off.
Is Claude Haiku 5.5 better than GPT-6 Luna?
On Anthropic’s own benchmarks, yes — Haiku 5.5 leads GPT-6 Luna on desktop operation (72.4% vs 48.9%), terminal tasks (39.2% vs 16.4%), coding (46.4% vs 42.4%), and office work ratings (1,620 vs 1,437). Both models cost the same per token. Independent benchmarks are still pending.
Should I use Claude Haiku 5.5 or Sonnet 5.5?
Use Haiku 5.5 for fast, high-volume work — subagents, summaries, classification, quick edits. Use Sonnet 5.5 (or Opus 5.5) for hard agentic coding and complex reasoning, where Anthropic itself says the bigger models win.
The Bottom Line on Claude Haiku 5.5
Claude Haiku 5.5 is a pricing event as much as a model release. At GPT-6 Luna’s price with materially better scores on computer-use and terminal benchmarks, it gives US developers and startups a cheaper default for the background AI work that used to eat budgets — provided your prompts stay under the 100K-token cliff. Watch for independent benchmarks before betting a production pipeline on Anthropic’s numbers, but for subagents, classification, and high-volume summarization, this is the small model to beat right now. That is Claude Haiku 5.5 explained: a $0.10 speed model for high-volume AI work — as long as your prompts stay under the 100K-token cliff.
Related reading: GPT-6 Intelligent UI Explained: What OpenAI’s October 2026 Update Changes for You · Gemini 4 Argon Explained: Google’s 1-Million-Token Answer to GPT-6 · Best AI Transcription Software 2026: 6 Tools Tested for Accuracy and Price
