Claude Haiku 5.5 pricing and features are the headline of Anthropic’s latest small-model release: a model that costs as little as $0.10 per million input tokens, handles a 1-million-token context window, and is the first Haiku model with an adjustable effort setting. Released on October 7, 2026, Haiku 5.5 is built for the work that fills most AI bills — summaries, classification, extraction, routing, and subagent tasks. Here is what it costs, what it can do, and how it compares with Haiku 4.5 and GPT-6 Luna.
Claude Haiku 5.5 Release Date and What It Is
The Claude Haiku 5.5 release date was October 7, 2026, and the model ID is claude-haiku-5-5. It sits below Sonnet 5.5 and Opus 5.5 as the fastest, cheapest option in the Claude 5.5 family, designed for high-volume, latency-sensitive tasks.
It is available on the Claude API, Claude apps, Claude Code, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
Claude Haiku 5.5 Pricing: The Two-Tier System Explained
Haiku 5.5 pricing splits at 100,000 prompt tokens — the single most important detail in the release. Rates per million tokens:
- Prompts up to 100,000 tokens: $0.10 input / $0.50 output
- Prompts over 100,000 tokens: $0.50 input / $2.50 output
- Prompt cache reads: $0.01 (up to 100K) / $0.05 (over 100K)
- Prompt cache writes: $0.125 (up to 100K) / $0.625 (over 100K)
Cache pricing is where high-volume workloads save the most: reused system prompts or document sets cost just $0.01 per million tokens on cache reads. On the Batch API, Haiku 5.5 supports up to 300,000 output tokens in beta, versus 128,000 on standard calls.
For scale, 10,000 short support conversations a day would cost roughly $1.70 at the under-100K tier — about ten times less than on Haiku 4.5.
Haiku 5.5 vs Haiku 4.5: Cheaper and Stronger
The Haiku 5.5 vs Haiku 4.5 comparison is striking on both price and capability. Haiku 4.5 charged $1.00 per million input tokens and $5.00 per million output tokens, so Haiku 5.5’s under-100K rates are up to 90% cheaper. Anthropic says Haiku 5.5 runs about 75% cheaper on average once the new tokenizer is accounted for (see the migration section below).
The capability jump is just as large, at least on Anthropic’s own benchmarks:
- OSWorld 2.1 (offline subset): 72.4% for Haiku 5.5, versus 15.7% for Haiku 4.5
- Terminal-Bench 4.0: 39.2% for Haiku 5.5, versus 0% for Haiku 4.5
- GDPval-AA v2.1: 1,620 for Haiku 5.5, versus 735 for Haiku 4.5
Two caveats: these are Anthropic’s figures, not independent test results, and the biggest gains are on computer-use and terminal-style tasks. For simple classification or summarization, the gain may be reliability and cost rather than a quality leap. Either way, Haiku 4.5 is now hard to justify for new work.
Claude Haiku 5.5 Pricing and Features: Comparison Table
| Feature | Claude Haiku 5.5 | Claude Haiku 4.5 | Claude Sonnet 5.5 |
|---|---|---|---|
| Input price (per million tokens) | $0.10 (up to 100K prompts); $0.50 (over 100K) | $1.00 | $2.00 |
| Output price (per million tokens) | $0.50 (up to 100K prompts); $2.50 (over 100K) | $5.00 | $10.00 |
| Prompt cache read (per million tokens) | $0.01 / $0.05 | Higher than Haiku 5.5 | $0.10 (halved at Haiku 5.5’s launch) |
| Context window | 1 million tokens | 200,000 tokens | 1 million tokens |
| Max output | 128,000 tokens (300,000 on Batch API beta) | Lower than Haiku 5.5 | 128,000 tokens |
| Adjustable effort setting | Yes — first Haiku with it | No | Yes |
| Best for | Summaries, classification, routing, subagents | Legacy small-model workloads | Complex agentic coding |
Sonnet 5.5 still costs 20 times more than Haiku 5.5 on input at the low tier, and Anthropic says it remains the better choice for complex agentic coding. Haiku 5.5 is not a Sonnet replacement; it is the model you hand the repetitive work to so Sonnet only handles the hard parts.
Claude Haiku 5.5 Context Window and Effort Setting
The Claude Haiku 5.5 context window is 1 million tokens — five times Haiku 4.5’s 200,000 — so you can feed in entire codebases or long documents without chunking them. But a prompt over 100,000 tokens pays the higher tier for the whole request, so the 1-million-token context is capacity, not a promise of the lowest price across all of it. Keep routine prompts under 100K, and cache long, repeated contexts.
The Haiku 5.5 effort setting is the model’s new control: the first Haiku with an adjustable effort parameter, trading cost and latency against answer quality per request. Adaptive thinking is on by default; default effort is medium. Use low effort for classification and routing, medium for everyday summarization and extraction, and higher effort when a task needs more reasoning. Test each level against your quality bar.
Haiku 5.5 vs GPT-6 Luna: The Price Match That Isn’t Quite Equal
The Haiku 5.5 vs GPT-6 Luna comparison is one buyers will make immediately, because the headline prices match: Luna lists the same $0.10/$0.50 short-context rates. The difference is where the higher tier kicks in — above 272,000 input tokens for Luna, versus above 100,000 for Haiku 5.5. For a 150,000-token prompt, Luna is cheaper on list price.
That does not settle which model is cheaper for a finished task: token consumption, retry rates, and accuracy all change the final bill. Benchmark both on your actual workload and compare cost per completed task, not cost per token. For short prompts the two are priced identically, so quality and latency should decide.
Best Uses for Claude Haiku 5.5
Haiku 5.5 shines wherever volume is high and each task is narrow:
- High-volume summarization — meeting notes, support tickets, and daily digests.
- Classification and extraction — tagging, sorting, and pulling structured fields from unstructured text.
- Routing — deciding which tool, team, or larger model should handle a request.
- Subagent work — the fast worker underneath a Sonnet or Opus model that plans and coordinates.
- Live support and voice agents — real-time conversations where latency is part of the product.
It is a poor fit for complex, multi-step agentic coding, where a wrong step is expensive and retries erase the savings.
Migration Gotchas: Read This Before Switching from Haiku 4.5
Two migration details, from Anthropic’s own guidance, can break a smooth upgrade:
- The new tokenizer counts more tokens. The same text counts roughly 30% more tokens than under Haiku 4.5’s tokenizer — already factored into Anthropic’s “about 75% cheaper on average” claim, but your token dashboards will jump even as your bill falls. Re-baseline budgets after migration.
- Non-default sampling settings return errors. A non-default
temperature,top_p, ortop_kreturns a 400 error on Haiku 5.5. Remove or reset those values before switching, and use the effort setting to control reasoning instead.
Run Haiku 5.5 in shadow mode alongside Haiku 4.5 before cutting over. If you need help planning that rollout, see our AI and automation services or contact us.
Frequently Asked Questions About Claude Haiku 5.5
When was Claude Haiku 5.5 released?
October 7, 2026. Its model ID is claude-haiku-5-5, and it is available on the Claude API, Claude apps, Claude Code, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
How much does Claude Haiku 5.5 cost?
Up to 100,000 tokens: $0.10 input / $0.50 output per million tokens. Above 100,000: $0.50 / $2.50. Cache reads cost $0.01 or $0.05, and cache writes $0.125 or $0.625, depending on the tier.
What is the Claude Haiku 5.5 context window?
1 million tokens, up from 200,000 on Haiku 4.5. Maximum output is 128,000 tokens, or up to 300,000 on the Batch API beta.
What is the Haiku 5.5 effort setting?
An adjustable control, new to the Haiku line, that trades cost and latency against answer quality per request. Adaptive thinking is on by default; default effort is medium.
Should I choose Haiku 5.5 or Sonnet 5.5?
Haiku 5.5 for high-volume, narrow tasks; Sonnet 5.5 ($2.00/$10.00 per million input/output tokens) for complex agentic coding and harder reasoning.
Conclusion: Claude Haiku 5.5 Pricing and Features Make Small Models Serious Again
Claude Haiku 5.5 pricing and features add up to a simple proposition: the cheapest Claude model is now far more capable, with a 1-million-token context window, an adjustable effort setting, and prices up to 90% below Haiku 4.5 for short prompts. Watch the 100,000-token pricing cliff and the new tokenizer, and leave the hardest work to Sonnet 5.5. For high-volume workloads, though, Haiku 5.5 turns jobs that were too expensive to run at scale into routine line items. Also see our comparison of ChatGPT vs GitHub Copilot vs Microsoft Copilot and our guide to using ChatGPT in web development.
Sources: MarkTechPost, TechRepublic, Pandromeda. Benchmarks are Anthropic’s own figures.
