Anthropic Releases Claude Haiku 5.5: 75% Lower Average Cost, Built for High-Frequency Repetitive Tasks and Partnering with Flagship Models
Faster and cheaper, Haiku 5.5 is tailored for high-volume summarization, search, and classification, or serving as a subagent alongside flagship models. Here is a breakdown of its performance gains, tiered pricing, adjustable effort, and early customer use cases.
When an AI application performs high volumes of summarization, lookup, and classification each day, the cost and latency of every model response add up quickly. Developers must balance two questions: can the model handle the task well, and is it affordable to run at scale over the long term?
Anthropic has released its lightweight model, Claude Haiku 5.5: its average running cost is about 75% lower than Haiku 4.5, alongside significant improvements in coding, computer use, and knowledge work. It is the fastest model in the Claude family at standard speeds. (Opus with Fast Mode enabled is faster.)
Its core capabilities and highlights center around these six points:
- Lower Cost: According to Anthropic’s estimate, average running costs drop by roughly 75% compared to Haiku 4.5, making it ideal for high-volume applications.
- Faster Response: The fastest in the Claude family under standard speed mode, built for latency-sensitive tasks like real-time customer support and browser navigation.
- Improved Task Performance: Shows substantial gains over Haiku 4.5 across official benchmarks in coding, computer use, and knowledge work.
- Adjustable Reasoning Effort: Introduces effort settings to the Haiku family for the first time, allowing lower resource spend on simple tasks and increased reasoning investment for difficult ones.
- Excels at High-Frequency Repetitive Work: Tailored for well-defined tasks such as summarization, context compression, database queries, and classification.
- Subagent Collaboration: Pairs with Opus 5.5 and Sonnet 5.5, letting flagship models direct complex operations while Haiku takes over specific routine steps.
In AI agent (autonomous systems that invoke tools and execute tasks step-by-step) applications, Opus 5.5 and Sonnet 5.5 can orchestrate complex workflows, while Haiku handles high-frequency subtasks like summarization, search, and categorization. This is the most practical advantage of this update: giving massive batches of small tasks a faster and far cheaper execution engine.
Short Request Rates Cut by 90%, Making High-Volume Invocations Far More Economical
When calling models via the API, inputs and outputs are billed by tokens (the fundamental units models use to process text). Haiku 5.5 splits pricing into two tiers based on the input length of a single request: for prompts up to 100,000 tokens, input and output unit prices are 90% lower than Haiku 4.5; for prompts exceeding 100,000 tokens, prices are 50% lower.
The rates below are listed in USD per 1 million tokens:
| Billing Item | Haiku 5.5 (≤100k Tokens) | Haiku 5.5 (>100k Tokens) | Haiku 4.5 | Sonnet 5.5 (Baseline Reference) |
|---|---|---|---|---|
| Input Price (Input) | $0.10 | $0.50 | $1.00 | $2.00 |
| Output Price (Output) | $0.50 | $2.50 | $5.00 | $10.00 |
| Cache Writes | $0.125 | $0.625 | $1.25 | $2.50 |
| Cache Reads | $0.01 | $0.05 | $0.10 | $0.10 |
In historical request data for Haiku 4.5, approximately 90% of requests had input lengths under 100,000 tokens, placing them into the new model's lowest pricing tier. For applications that frequently perform summarization, classification, and searches, the savings per invocation compound significantly across millions of calls.
The roughly 75% reduction in average running costs is Anthropic's composite estimate factoring in both pricing tiers and empirical task usage. Because Haiku 5.5 incorporates an updated tokenizer, the same task consumes slightly more tokens, which accounts for the difference between the nominal unit price reduction and actual workload cost reductions.
When reusing system prompts or background context, caching allows identical input prefixes to be reused at reduced rates. Cache reads for short requests cost just $0.01 per million tokens, making it ideal for workflows that repeatedly reference the same background material.
Calculate the Cost Difference for the Same Workload
Compare based on a baseline of 10,000 input tokens, 1,000 output tokens, and 1,000 runs. Adjust the input length to see the rate changes above 100k tokens.
Prompt ≤100k tokens: Input $0.10, Output $0.50 / MTok.
Cost = (Input tokens × Input price + Output tokens × Output price) ÷ 1M × Runs. This comparison assumes equal token volume without caching; actual token consumption, retries, and tool calls will affect final expenses.
Where Do the Small Model's Gains Specifically Show Up?
The improvements in Haiku 5.5 span knowledge work, computer use, coding, and chart comprehension. In benchmarks released by Anthropic, scores across all these domains show significant leaps over Haiku 4.5.
Key benchmark results are summarized below:
| Benchmark & Domain | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 (Reference) |
|---|---|---|---|---|
| OSWorld 2.1 Offline Subset (Computer use test; offline subset scored on partial credit) | 72.4% | 15.7% | 48.9% | 83.9% |
| GDPval-AA v2.1 (Professional knowledge work across 44 occupations, Elo rating) | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 (Comprehensive business and professional case evaluations, Elo rating) | 1578 | 614 | 1336 | 1824 |
| Humanity’s Last Exam (No Tools) (Multidisciplinary frontier academic problems and deep reasoning) | 45.9% | 10.2% | — | 56.9% |
| Humanity’s Last Exam (With Tools) | 57.4% | 18.7% | — | 64.5% |
| Terminal-Bench 4.0 (Agent benchmark for autonomous debugging and coding in a CLI terminal) | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 Main Track (Frontier coding capabilities) | 46.4% | — | 42.4% | 52.1% (Xhigh effort) |
| Chartography (No Tools) (Complex chart analysis and visual reasoning) | 46.4% | 6.4% | 29.1% | 61.6% |
The progress in computer use and professional knowledge work is particularly striking: OSWorld offline subset scores surged from 15.7% to 72.4%, and GDPval Elo jumped from 735 to 1620. Chart comprehension also advanced substantially, with tool-free Chartography scores climbing from 6.4% to 46.4%. These improvements equip Haiku to handle concrete steps in document analysis, computer interaction, and targeted coding.
Complex programming remains best suited for flagship models: on Terminal-Bench 4.0, Haiku 5.5 scored 39.2% compared to Sonnet 5.5's 70.6%. Deploying Haiku for well-scoped subtasks alongside Sonnet or Opus for overall workflow coordination aligns directly with its intended role.
Spend Less Thinking on Simple Tasks, Invest More on Hard Ones
Haiku 5.5 introduces adjustable effort settings to the Haiku family for the first time, giving developers control over how much reasoning capacity the model deploys. You can select lower effort for rapid classification, or ramp up effort for more challenging document analysis and computer use to achieve stronger performance at higher cost.
Three evaluation charts illustrate performance changes across five effort tiers: Low, Med, High, Xhigh, and Max. (The horizontal axis uses a logarithmic scale for USD cost, with expenses increasing to the right; curve readings below represent approximations.)
Computer Use: Even low effort markedly surpasses the previous generation. Across the five tiers, Haiku 5.5 scores roughly 42%, 53%, 61%, 68%, and 72%, with cost per attempt increasing from about $0.07 to $0.62. The lowest effort tier already outperforms Haiku 4.5's ~16% while costing less. The Sonnet curve tops out at around 84%.
Professional Knowledge Work: Increasing investment steadily improves results. Haiku's Elo rating rises from around 1120 in the Low tier to approximately 1620 in the Max tier, with cost per task growing from about $0.012 to $0.85. Sonnet's reported score range in the official blog spans 1179 to 1840, indicating that reasoning effort substantially shapes cross-model comparisons.
Academic Frontier Problems: Diminishing returns appear at the highest tiers. On tool-free HLE, Haiku's score rises from roughly 31% at Low to around 46% at Max, while cost per attempt increases from approximately $0.004 to $0.225; gains from Xhigh to Max narrow visibly. The highest score cited for Sonnet on the official blog is 56.9%.
For high-frequency tasks, the value of effort lies in allocating budgets by problem difficulty: resolve simple requests as swiftly as possible, and allocate more reasoning to intricate ones. The tapering curve at higher tiers also demonstrates that defaulting to Max effort is not always cost-effective.
Flagship Models Drive the Workflow, Haiku Handles Specific Steps
A complete task often consists of many small constituent steps. Through multi-agent collaboration, a flagship model can offload search, extraction, and summarization subtasks to Haiku subagents (auxiliary agents called by the primary agent to execute defined tasks), significantly lowering latency and cost for those steps.
Early customer testing highlighted in the official blog demonstrates several concrete use cases:
- Financial Presentation Assistance (Rogo): When building presentation decks, a larger model directs the overall narrative structure, while Haiku 5.5 acts as a subagent to extract specific segment revenue line items from 10-K filings (annual reports filed by public companies with US regulators). Its precision is well suited for targeted extraction, delivered at high speed and low cost across frequent calls.
- Autonomous Coding Collaboration (Cognition / Devin Fusion): Within the Devin Fusion architecture, Haiku 5.5 acts as a sidekick paired with Opus 5.5 as the lead model. This combination preserved a benchmark score of 66.2 on FrontierCode while reducing invocation cost and latency. Developers can already test this Opus 5.5 lead and Haiku 5.5 assistant setup in Devin CLI.
- Enterprise Collaboration and Task Routing (Asana): In evaluations for AI Teammates, Haiku 5.5 was deployed to classify tickets, initialize projects, and detect high-risk and overdue tasks. Tests demonstrated a greater than 30% reduction in end-to-end task completion latency, with individual agent turn inference speeding up by up to 2.5x.
- High-Concurrency Document Q&A (AlphaSense): Its production "Ask in Document" workflow processes roughly 8 million queries per week, answering targeted questions anchored to single or few documents. In an evaluation across 400 queries, Haiku 5.5 improved test scores from 0.76 with Haiku 4.5 to 0.84, representing a statistically significant gain.
- CRM Data Auditing (HubSpot) (CRM stands for Customer Relationship Management): In simulated CRM task benchmarks, Haiku 5.5 achieved an average test score of 92.8% across three runs. On an auditing task identifying stale and ambiguous CRM records, Haiku 5.5 was the fastest model tested while delivering the lowest false positive rate and highest hit rate.
- Enterprise Document Analysis (Box): In early testing, Haiku 5.5 scored 11 points higher than Haiku 4.5 with roughly half the latency. Box plans to deploy it for large-scale analysis tasks including expense reports, financial summaries, and weekly business reviews.
Which Workloads Are Best Suited for Haiku?
- Repetitive Summarization and Context Compression: Condensing accumulated context within long workflows to prepare concise backgrounds for downstream steps.
- Targeted Queries, Extraction, and Categorization: Looking up single data points in documents, executing database queries, or classifying support tickets and records.
- Interactions Requiring Rapid Responses: Real-time customer support, browser manipulation, and scoped subtasks dispatched by a lead agent.
These tasks share clear objectives and high invocation volumes, meaning improvements in unit cost and response speed quickly compound into substantial real-world gains.
What Safety Guardrails and Ecosystem Changes Accompany the Release?
Safety policies have also been updated alongside model capabilities:
- Value Alignment: Anthropic reports that Haiku 5.5 demonstrates significant improvements over Haiku 4.5 across nearly all internal alignment evaluations, showing fewer deviations from instructions and lower willingness to comply with abusive requests.
- Cybersecurity Safeguards: Defensive restrictions are stricter than Haiku 4.5, yet allow a broader range of defensive analysis tasks compared to Sonnet 5.5, while continuing to restrict dual-use techniques like penetration testing that are prone to exploitation.
- Biological Safety: Biosecurity safeguards are aligned with Sonnet 5, Sonnet 5.5, and Opus 5. Standard bioscience research inquiries are permitted, while requests assessed as posing potential hazards remain restricted. Organizations needing broader access for biological or cyber research can apply for life sciences or cyber verification programs.
Alongside this launch, Anthropic introduced three accompanying updates:
- Sonnet 5.5 Cache Read Price Cut by 50%: Pricing drops from $0.20 to $0.10 per million tokens. Long-running agents repeatedly read the same context, and cache reads account for a large portion of tokens; Anthropic estimates this price drop reduces total operating costs for most agent workloads by roughly 20%.
- Monthly API Credits for Subscribers: Rolling out this week to Claude Max and Team subscribers as dedicated monthly API credits on the Claude Platform: Max 5x subscribers receive $100 per month, Max 20x subscribers receive $200 per month, and Team accounts receive up to $500 per month as a shared pool across members. These credits can be applied toward calling any model on the Claude Platform to prototype tools, applications, and agents. Further details are available in the official Help Center article.
- Official SDK Support for Computer and Browser Use (Beta): The Python and TypeScript SDKs have been updated concurrently to introduce beta support for computer use and browser use.
Claude Haiku 5.5 is currently available on the Claude Platform (under the model ID claude-haiku-5-5) as well as across Amazon AWS, Google Cloud, and Microsoft Azure. For integration steps and migration details, see the Haiku 5.5 Migration Guide Explainer.
Members can copy or download the full Markdown, and let their AI subscribe and ask.
Found this useful? Leave your email — every new explainer, the moment it's out. Unsubscribe anytime.