Claude Haiku 5.5 Explained: Pricing, Benchmarks, API Setup and the 90% Cost-Cut Claim

Anthropic officially launched Claude Haiku 5.5 on October 7, 2026, releasing its cheapest and fastest small Claude model to date. For API requests using prompts of no more than 100,000 tokens, Anthropic charges $0.10 per million input tokens and $0.50 per million output tokens. It is designed for high-volume support, document extraction, classification, compact summaries, coding subagents and fast browser interactions.

Is Haiku 5.5 really 90% cheaper? The listed token prices are 90% lower than Claude Haiku 4.5 at that short-prompt tier. However, Anthropic estimates that typical workloads cost about 75% less because the newer model may use more tokens, including an approximately 30% increase in token counts for comparable text under its new tokenizer. More importantly, requests exceeding 100,000 prompt tokens cost five times the short-prompt rate. Developers need to understand that threshold before scaling.

This independently explained guide brings together Anthropic's official launch announcement, model specifications, benchmark tables, pricing, availability and migration documentation. It includes a developer tutorial, a cost calculator, scenario comparisons and answers to common reader questions. The pricing and capabilities were checked October 15, 2026.

Claude Haiku 5.5 launched October 7, 2026. Its official model ID on the Anthropic Claude API is claude-haiku-5-5. On Amazon Bedrock, Anthropic lists anthropic.claude-haiku-5-5. Google Cloud, Microsoft Foundry and the Claude Platform on AWS are also listed among supported deployment options, but platform access, region availability and billing terms can differ.

Haiku 5.5 completes the 5.5 Claude model family, following Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Anthropic says Free, Pro, Max, Team and Enterprise Claude app users can select Haiku 5.5 on supported web and mobile interfaces. The model is also accessible within Claude Code.

It is important to distinguish product access from metered developer usage. Selecting Haiku in a consumer application may be available with plan-specific limits. Calling the separate API uses metered tokens and ordinarily requires an account configured for API usage, except where qualifying credits apply.

Haiku is the speed-and-cost tier of Anthropic's Claude lineup. Anthropic positions 5.5 for repetitive, narrowly defined work that does not justify the cost of a frontier model on every request.

A customer-support system might use Haiku to identify a customer's problem, retrieve a relevant company-policy excerpt and draft a grounded answer. A finance application might ask it to extract a revenue figure from an annual report. A coding agent might delegate searches, summaries, narrow changes and tests to Haiku while a larger model plans the project.

Other applications include content moderation assistance, document compaction, CRM classification, ticket routing, structured extraction, browser form-filling, email categorization and short-form communication. Those are potential use cases, not guarantees of perfect results or blanket approval for sensitive automated decisions.

The deployment question is how frequently the task is performed, how much context is sent, how difficult the job is and what happens if the model makes a mistake. A small model can be economical for one hundred thousand routine tasks and inappropriate for one unchecked high-stakes decision.

Anthropic offers two pricing tiers determined by the complete prompt length of each API request. For prompts up to and including 100,000 input tokens, Haiku 5.5 is $0.10 per million input tokens and $0.50 per million output tokens. For prompts over 100,000 tokens, the rate is $0.50 input and $2.50 output.

This is not a progressive rate in which only tokens beyond 100,000 are expensive. A request with a 150,000-token prompt is charged at the higher input and output prices for that request. The prompt-length test includes cached and uncached prompt tokens. Prompts already billed on previous requests are not repriced retroactively.

The official basic 5-minute cache write rate is $0.125 per million tokens at the short-prompt tier; the 1-hour cache write rate is $0.20, and cache reads cost $0.01. At the long-prompt tier, the respective amounts are $0.625, $1.00 and $0.05. The minimum cacheable prompt is now 512 tokens rather than Haiku 4.5's 4,096-token minimum.

Eligible asynchronous Batch API requests receive a 50% discount on ordinary input and output token costs. Caching and batch processing can be valuable for offline processing, but a real-time voice agent generally cannot wait for asynchronous batch completion.

The interactive cost calculator in this report shows estimates under the standard first-party pricing structure, excluding cloud-platform markups, auxiliary tools, data-residency surcharges, taxes and retries.

Consider an application whose average request uses 10,000 input tokens and 1,000 output tokens. At the short-prompt rate, the model-only cost is $0.001 for input plus $0.0005 for output, totaling $0.0015 per call.

At constant token counts, that is $1.50 for 1,000 calls, $150 for 100,000 calls and $1,500 for one million calls. The equivalent listed Haiku 4.5 prices of $1 per million input and $5 per million output would calculate to $1,500 per 100,000 calls, using identical token counts.

These are transparent arithmetic examples, not a quote for every customer's production system. The newer model may count more tokens for the same text, responses can vary in length, and optional tools, retries and fallbacks add cost. A more useful business KPI is cost per successfully resolved task rather than cost per unverified model response.

A long-context request with 150,000 input tokens and 4,000 output tokens would instead cost about $0.085 at the higher tier: $0.075 input and $0.010 output, excluding other fees. This demonstrates why aggressive conversation trimming and document selection can materially change the business case.

Anthropic's published benchmark comparison shows substantial gains over its preceding small model. On OSWorld 2.1's specified offline subset, Haiku 5.5 scored 72.4% compared with 15.7% for Haiku 4.5. Sonnet 5.5 scored 83.9%. OSWorld measures computer-using agents in a controlled evaluation, not universal success on real business websites.

On Humanity's Last Exam without tools, Haiku 5.5 scored 45.9%, up from 10.2% for Haiku 4.5; Sonnet 5.5 scored 56.9%. With tools, Haiku 5.5 scored 57.4%, versus 18.7% for Haiku 4.5 and 64.5% for Sonnet 5.5.

On the demanding agentic coding evaluation Terminal-Bench 4.0, Haiku 5.5 reached 39.2%, compared with 0.0% for Haiku 4.5 and 70.6% for Sonnet 5.5. That gap is an important reminder that the cheaper model is not the strongest option for complicated multistep coding.

Anthropic also reported a 1,620 GDPval-AA v2.1 score for Haiku 5.5, against 735 for Haiku 4.5 and 1,840 for Sonnet 5.5. This benchmark uses a different scale; its figures cannot be read as percentages.

RecoupRev did not rerun these benchmarks independently. They are attributed to Anthropic's official October 7 release under the benchmark conditions described there. Model choice should also account for real workload error rates, latency and tool usage.

Anthropic describes Haiku 5.5 as its fastest model at standard speed. In testing reported by early customer Asana, task completion latency fell by more than 30% and inference per agent turn improved by up to 2.5 times against Asana's comparison model. Another customer, HubSpot, reported a 92.8% averaged score on its simulated CRM task suite.

These are attributed customer observations, not independently measured universal performance claims. Your latency will depend on output size, reasoning effort, network conditions, concurrency, rate limits, backend tools and the nature of the task. The larger Claude models may run better on specific complex tasks even when their raw first-token latency is higher.

For a production comparison, measure p50 and p95 response time, time to first token, the share of tasks completed correctly, number of tool calls and dollar cost for each valid result. A fast answer that repeatedly needs human correction may be more expensive than a slower, more reliable alternative.

Haiku 5.5 supports a one-million-token context window and a standard maximum output of 128,000 tokens, compared with Haiku 4.5's 200,000 context and 64,000 maximum output. This greatly expands the size of documents and conversations that the model can technically process.

However, the high context ceiling does not make a 900,000-token request economical. Haiku 5.5 applies the higher price tier whenever the prompt exceeds 100,000 tokens, even if some of the content was retrieved from cache. Nor does a large context window guarantee accurate recall of every detail buried in a long document.

For many applications, a good retrieval pipeline that selects a few relevant pages is more cost-effective than repeatedly sending a complete document repository. Consider chunking, citations, search, conversation compaction and verification whenever the data set grows.

There may also be separate message-size, platform, rate-limit and image/PDF restrictions. The model's technical context limit should not be presented as an unlimited account entitlement.

According to Anthropic's Haiku product page, Haiku 5.5 can be selected by users of Claude's Free and paid application tiers where available. Sign in to the Claude app, open the model selector and choose Haiku 5.5. Usage and feature restrictions may depend on your current plan.

This is different from API pricing. A free Claude chat account is not a license to operate an unlimited commercial API or to share personal credentials with customers. For a hosted app, use your own server-side API credentials and a legitimate billing arrangement.

Anthropic also announced monthly API credits for qualifying Max and Team customers: Max 5x users receive $100 monthly, Max 20x users $200, and Team customers up to $500 pooled across users, subject to the applicable rollout and connection requirements. These do not mean every user gets an unconditional monthly API allowance.

Start by creating a Claude Platform account and generating an API key. Install an up-to-date Python environment and place the key in an environment variable, not inside your application's source file or version-control history. On Linux, macOS or Windows Subsystem for Linux, the basic setup is:

Create haiku_demo.py and use the following script to send a test message:

Run it using python haiku_demo.py. The script prints text blocks from the returned message. A successful call shows that the account can access the model and that the API environment variable is available.

Usage is metered. Do not expose a production key in a browser app, publish it in a repository, include it in error reports, or share it with untrusted third-party API resellers.

Developers who do not want to install the Python SDK can call the Messages API directly. This example uses the same model ID, an API key stored in the environment and Anthropic's API version header:

A response contains message content and token-usage information. It should not be treated as validation of an application's production correctness; introduce representative tests, timeouts, retry logic and billable-usage tracking before scaling.

Claude Haiku 5.5 is the first Haiku-class model with adjustable effort. Anthropic recommends low effort for fast, predictable requests, medium as the default for many tasks and high for harder knowledge or tool-use work. The documentation also describes xhigh and max effort settings when extensive reasoning demonstrably improves outcomes.

Adaptive thinking is on by default. Thinking can consume part of max_tokens, so a response limit that is too small may allow reasoning but little visible text. The old manual budget_tokens configuration is a breaking migration case: developers should use the current adaptive-thinking API and official documentation rather than simply changing model IDs and reusing all settings.

For cost-sensitive production apps, start with medium, test low, then measure whether higher effort improves the fraction of correct answers. A higher token count is not automatically more useful if the task only needs a label or short answer.

The model's official migration guide lists changes from Haiku 4.5 that developers should test. Legacy manual extended thinking with budget_tokens can cause errors. Some sampling parameters, including temperature, top_p and top_k, are no longer supported in the same way. Older computer-use tool IDs need updating and the required replacements can differ across the first-party API and cloud providers.

Assistant-message prefill is also not supported in the previous form. The model does not inject some old context-awareness tags automatically. If an agent needs task-level limits, implement explicit application budgets and consult Anthropic's task-budget documentation.

The tokenizer change is commercially significant. Anthropic says equivalent text may generate approximately 30% more tokens than on Haiku 4.5, depending on content. Re-evaluate max_tokens, token-based pricing estimates, rate limits and cache-hit behavior before migrating a production app.

A safe release process runs real representative tasks against both models, checks accuracy and refusals, measures cost, and gradually shifts traffic while retaining a route to recover from errors.

Haiku 5.5 is most compelling for classification, extraction, summarization, quick replies, request routing, simple code edits, browser assistance and narrowly scoped subagents. Sonnet 5.5 offers more capability on difficult coding and knowledge work, while Opus 5.5 targets complex, sustained tasks where greater judgment may justify higher cost.

For short prompts, the official listed input/output rates are Haiku 5.5 at $0.10/$0.50 per million tokens; Sonnet 5.5 at $2/$10; and Opus 5.5 at $4/$20. These price differences are substantial, but picking a model only by token rate can backfire if the cheaper model requires many retries or escalations.

An effective system can let a more capable model handle planning and unusual exceptions while Haiku handles repeatable substeps. The optimal share of Haiku traffic is something to determine from real quality tests, not a fixed universal rule.

Haiku 5.5's low latency and cost make it relevant to voice and customer-message systems. For a conventional voice pipeline, speech-to-text converts audio into text; the language model produces an answer or tool request; a text-to-speech provider converts the response back to audio. The total response time also includes transport, turn detection, database lookups and synthesis.

Haiku can help with appointment qualification, customer intents, support-ticket summaries, policy-grounded answers and narrow structured lookups. For WhatsApp bots, it can classify a request and draft a reply, but the bot still requires a compliant messaging integration, backend logic and permission checks.

In healthcare, financial services or other high-stakes support, the model should not be left to authorize transactions, change records or give consequential advice without additional controls. Confirm that every action is permitted, use source-grounded outputs and route uncertain cases to appropriate review.

Anthropic reports large gains on a computer-use benchmark and says its Python and TypeScript SDKs are being updated for beta browser and computer-use tools. These can be useful for repetitive desktop or web tasks, including navigation and data entry.

The model cannot safely be treated as a universal autonomous employee. Real applications have confidential data, credential boundaries, website policies and irreversible actions. Use allowlisted tools, approval before consequential changes, audit logging and a clear task boundary.

As a coding subagent, Haiku can search files, summarize logs or make simple targeted edits. Sonnet and Opus can coordinate harder projects, but automated tests and a review process remain important for security-sensitive changes.

Anthropic reports improved alignment performance for Haiku 5.5 and more restrictive cybersecurity safeguards than on Haiku 4.5. Its safety systems may refuse some requests; apps should handle stop_reason set to refusal rather than repeatedly resubmitting the same prompt.

Prompt injection remains a risk when models read webpages, documents, emails and tickets containing untrusted instructions. Treat such data as reference material, not commands. Do not allow a model to expose keys, follow embedded malicious instructions or invoke powerful tools without independent authorization.

Enterprise reliability requires measurement. Test false positives, missing fields, incorrect citations, tool-call validation and the frequency of needed human interventions. Benchmarks from Anthropic and early partners can guide model selection, but they are not a substitute for evaluation on your own documents and user behavior.

Search intent around the October release includes the exact model name, pricing, API access, free usage, benchmarking, long context, Python installation and comparisons with Sonnet 5.5. Relevant phrases for this guide include Claude Haiku 5.5 pricing; Claude Haiku 5.5 API; Haiku 5.5 vs Haiku 4.5; Haiku 5.5 vs Sonnet 5.5; claude-haiku-5-5 model ID; Claude Haiku 5.5 benchmark; how to use Claude Haiku 5.5; and Claude Haiku 5.5 Python example.

These terms are relevant to questions potential readers may type into search engines. A public Google Trends search alone does not establish independent rankings, search volumes or breakout increases for every phrase, and RecoupRev does not invent those figures. For a developing technology story, accurate answers, sources and dates matter more than repetitive keyword placement.

Yes. Anthropic's release notes and announcement give October 7, 2026 as the release date.

At most 100,000 prompt tokens: $0.10 per million input tokens and $0.50 per million output tokens. Above 100,000 prompt tokens: $0.50 input and $2.50 output per million tokens for the entire request.

Yes. Its standard maximum output is 128,000 tokens. Long-prompt pricing and independent request-size limits still apply.

Anthropic lists it as selectable on the Free tier, subject to app access and usage limits. This is not an unlimited free API.

claude-haiku-5-5 on Anthropic's first-party API. On Amazon Bedrock, Anthropic lists anthropic.claude-haiku-5-5.

Anthropic reports much stronger benchmark results and lower short-prompt token prices. Your effective gain depends on the workload, prompts, tokenization, tool use and quality requirements.

Anthropic markets Haiku 5.5 as its fastest model at standard speed. Actual end-to-end latency varies with output length, effort, tools and deployment conditions.

Yes, Anthropic lists it as available in Claude Code. Access and usage limits depend on the account and product.

At 10,000 input tokens and 1,000 output tokens per call, approximately $150 in basic model fees at the lower tier, excluding tools, retries, caches and other service costs.

No. The higher pricing rate applies to the request when its prompt exceeds 100,000 tokens, not only to the portion above the threshold.

Yes, review the migration guide. Manual thinking budgets, sampling parameters, assistant prefills and some computer tools have changed.

Anthropic's published Terminal-Bench 4.0 figures favor Sonnet 5.5 over Haiku 5.5 on complicated agentic coding. For narrow, high-volume jobs, Haiku may be better suited on cost and latency.

Claude Haiku 5.5 makes the most sense for teams that can define success clearly and run large quantities of short, repetitive AI work. Its speed, low short-prompt prices, updated benchmark results, large context option and adaptive reasoning settings are compelling capabilities for production automation.

The pricing catch is real: prompts over 100,000 tokens use higher rates, and actual workload token counts can rise under the new tokenizer. Its coding benchmarks also show it has not displaced the larger models on the hardest tasks. The rational implementation is to measure cost per successful result, trim context, test both effort levels and escalate harder tasks appropriately.

Research date: October 15, 2026. Sources: Anthropic October 7 announcement; official Claude model, pricing, migration, prompting, and context documentation; and attributed customer evaluations. RecoupRev has not independently reproduced the provider's benchmark suite.