Claude Haiku 5.5 Compact Model: What Changes

A New Benchmark for Small Models

A procurement team runs 40,000 document-classification requests every week. Each one hits an AI model: extract the vendor name, flag the spend category, route to the right approval queue. The model is fast, the logic is simple — but at scale, the token bill quietly becomes a line item that finance starts asking about. On October 7, 2026, Anthropic released Claude Haiku 5.5, and that line item just got a lot smaller.

The price drop is real and significant, but it's only part of the story. What Anthropic shipped alongside the new pricing is a capability profile that looks nothing like a typical "budget" model — and the implications for how businesses structure their AI agent stacks are worth examining carefully before you assume this is just a cheaper version of what you already have.

The Claude Haiku 5.5 compact model replaces Haiku 4.5 as Anthropic's fastest and most affordable offering. It sits at the base of the Claude 5.5 family — below Sonnet 5.5 and Opus 5.5 — and is designed for the high-volume, latency-sensitive work that makes enterprise AI expensive at scale: document summarization, information classification, database queries, and acting as a sub-agent inside larger orchestration pipelines.

The Numbers That Actually Matter

Pricing: Two Tiers, One Big Shift

Haiku 5.5 introduces a two-tier pricing structure split at the 100,000-token prompt boundary. For prompts under that threshold — which covers the vast majority of typical business requests — the model costs $0.10 per million input tokens and $0.50 per million output tokens. That's a 90% reduction per token compared to Haiku 4.5, which was priced at $1.00 input and $5.00 output per million tokens.

For longer prompts above 100,000 tokens, pricing rises to $0.50 input and $2.50 output — still a 50% cut versus the predecessor. Anthropic estimates that across realistic mixed workloads, the average running cost comes out roughly 75% lower than Haiku 4.5. According to Anthropic's own data, approximately nine in ten Haiku 4.5 requests fall under the 100K-token line, so most production deployments will see savings close to the 90% figure.

One practical caveat: Haiku 5.5 uses a newer tokenizer that converts identical text into roughly 30% more tokens than the previous version. Factor that into your migration math before projecting savings — the net reduction is still substantial, but the raw per-token discount overstates the real-world gain slightly.

The gap between "cheap model" and "capable model" used to be a trade-off businesses accepted. With Haiku 5.5, Anthropic is arguing that gap has closed — at least for the category of work that drives most of the volume.

Benchmark Performance: The OSWorld Story

The pricing headline is attention-grabbing, but the benchmark data is where the strategic picture sharpens. On OSWorld 2.1 — a test that measures a model's ability to complete tasks by operating a real desktop environment, navigating interfaces, clicking through menus, and handling browser-based workflows — Haiku 5.5 scores 72.4% on the offline subset. Haiku 4.5 scored 15.7% on the same test.

That's not an incremental improvement. It's a 56-point jump that puts Haiku 5.5 ahead of OpenAI's GPT-6 Luna (48.9% on the same benchmark) at identical pricing of $0.10 per million input tokens. Sonnet 5.5 still leads at 83.9%, but it costs twenty times more per token at the standard tier.

On Terminal-Bench 4.0, which tests agentic coding in terminal environments, Haiku 4.5 scored 0.0% — a complete failure on the benchmark. Haiku 5.5 reaches 39.2%. On Humanity's Last Exam with tools enabled, the score moves from 18.7% to 57.4%.

The pattern across all benchmarks is consistent: Haiku 5.5 is not a marginal upgrade. It's a model that can now handle structured agentic tasks that its predecessor simply couldn't.

What "Adaptive Reasoning" Means in Practice

One of the less-discussed features of Haiku 5.5 is its adaptive thinking capability — the model's ability to scale its reasoning depth based on task complexity. This makes it the first Haiku model to support effort levels, similar to what Sonnet and Opus have offered.

In practical terms: a simple classification task gets a fast, shallow response. A multi-step browser automation sequence gets deeper reasoning before the model acts. The model calibrates rather than applying maximum compute to every request regardless of need.

For businesses running mixed workloads through a single model endpoint, this matters. You're not paying for heavy reasoning on tasks that don't require it, and you're not getting shallow responses on tasks that do. The default effort level is set to medium, which Anthropic positions as the right balance for most production use cases.

This is also what makes Haiku 5.5 a natural fit for multi-agent architectures where a lightweight model handles routing, triage, and sub-tasks while a more capable model handles the exceptions. The economics of that architecture just improved significantly.

Where Businesses Should Actually Deploy This

High-Volume Document Processing

The clearest immediate use case is any workflow that processes large numbers of short-to-medium documents: invoice classification, contract metadata extraction, support ticket routing, compliance flag detection. At $0.10 per million input tokens, running 10 million tokens of document processing costs $1. That's not a rounding error — it's a structural change in what's economically viable to automate.

Teams that previously sampled documents for AI review because full coverage was too expensive can now run complete coverage. That changes the compliance and audit story considerably.

Computer Use and Browser Automation

The OSWorld improvement is the most strategically significant capability shift. A model that scores 72.4% on autonomous desktop task completion — navigating real interfaces, not just generating text — opens up a different category of automation: the kind that previously required RPA tools or custom scripting.

Think of the workflows that live in legacy web portals, internal dashboards, or vendor systems that don't offer APIs. Haiku 5.5 can navigate those environments at a cost point that makes it viable to run continuously, not just as an occasional experiment. For a deeper look at how Anthropic's approach to agent control shapes these deployments, the analysis in Anthropic Can't Control AI Agents: What It Means is worth reading alongside this.

Sub-Agent Roles in Larger Pipelines

In orchestrated AI systems, the expensive model shouldn't be doing everything. Haiku 5.5 is built for the roles that generate most of the token volume: parsing inputs, formatting outputs, making routing decisions, handling retries, summarizing intermediate results. Reserving Sonnet or Opus for the genuinely hard reasoning steps — and routing everything else through Haiku 5.5 — is now the economically rational architecture.

When the cost of a sub-agent drops by 75-90%, the question stops being "can we afford to automate this?" and becomes "what's the cost of not automating it?"

Availability and Integration

Haiku 5.5 is available immediately across Anthropic's own platform, Amazon Bedrock, Google Cloud, and Microsoft Azure, under the model identifier claude-haiku-5-5. Python and TypeScript SDK updates add beta support for browser and computer use capabilities directly.

The model carries a 1-million-token context window with a 128K maximum output — specs that are consistent with Sonnet 5.5 and appropriate for the document-heavy workflows where Haiku 5.5 will see the most use.

For teams already running Haiku 4.5 in production, the migration path is straightforward at the API level. The main adjustment is accounting for the tokenizer change and the two-tier pricing structure when estimating costs on longer prompts.

The Competitive Context

Haiku 5.5's pricing matches OpenAI's GPT-6 Luna at $0.10 per million input tokens — and on Anthropic's published benchmarks, it outperforms Luna on every listed test. OSWorld: 72.4% versus 48.9%. Terminal-Bench 4.0: 39.2% versus 16.4%. FrontierCode 1.1: 46.4% versus 42.4%.

This is the context in which Anthropic describes Haiku 5.5 as "the cheapest, fastest, and most capable small model we've ever released." The claim holds up against the published data. Whether it holds in your specific workload depends on your prompt length distribution and task mix — but the benchmarks give a credible baseline.

The broader trend this reflects is worth naming: the cost of capable AI inference is falling faster than most enterprise AI budgets anticipated. Organizations that locked in cost assumptions for AI automation twelve months ago are working with numbers that are increasingly out of date. The AI ROI framework question isn't just "does this pay off?" — it's "does our cost model still reflect what the technology actually costs?"

FAQ

What is Claude Haiku 5.5 and when was it released? Claude Haiku 5.5 is Anthropic's fastest and most affordable small language model, released on October 7, 2026. It replaces Haiku 4.5 as the entry-level model in the Claude 5.5 family and is designed for high-volume, latency-sensitive tasks including document processing, classification, and computer use automation.

How much does Claude Haiku 5.5 cost compared to Haiku 4.5? For prompts under 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens — roughly 90% less per token than Haiku 4.5's $1.00/$5.00 pricing. Anthropic estimates the average running cost across realistic workloads is about 75% lower than its predecessor, accounting for the new tokenizer and prompt length distribution.

What is OSWorld and why does Haiku 5.5's score matter? OSWorld 2.1 is a benchmark that tests whether an AI model can complete tasks by operating a real desktop environment — navigating interfaces, clicking through menus, and handling browser-based workflows autonomously. Haiku 5.5 scores 72.4% on the offline subset, up from 15.7% for Haiku 4.5. This jump signals that the model can now handle structured computer-use automation that its predecessor couldn't, at a cost point that makes continuous deployment viable.

What is adaptive reasoning in Haiku 5.5? Haiku 5.5 is the first Haiku model to support effort levels — it scales its reasoning depth based on task complexity. Simple tasks get fast, shallow responses; complex multi-step tasks get deeper reasoning before the model acts. The default is set to medium effort, which Anthropic positions as appropriate for most production workloads.

Where is Claude Haiku 5.5 available? The model is available on Anthropic's own platform, Amazon Bedrock, Google Cloud, and Microsoft Azure under the identifier claude-haiku-5-5. Python and TypeScript SDKs include beta support for computer use and browser automation capabilities.

Should I migrate from Haiku 4.5 immediately? For most high-volume workloads under 100,000 tokens per request, the economics strongly favor migration. The main consideration is the new tokenizer, which generates roughly 30% more tokens from identical text — meaning the effective savings are real but slightly less than the raw per-token discount suggests. Test your actual prompt distribution before finalizing cost projections.

---

The release of Haiku 5.5 is one of those moments where the economics of a technology shift enough to change the decision calculus — not just for teams already running AI automation, but for the ones that were waiting for the cost curve to make sense. A model that can navigate a desktop, classify documents, and act as a capable sub-agent in a multi-step pipeline, at $0.10 per million input tokens, removes the "too expensive at scale" objection from a wide range of automation projects.

For executives who've been watching AI costs carefully, this is the kind of development that turns a pilot into a production rollout. And for the board or investors asking whether the company is building AI infrastructure that actually scales — having a clear answer to that question, backed by concrete numbers rather than vague commitments, is exactly the kind of confidence that separates leaders who are ahead of the curve from those still catching up.

If you're evaluating where Haiku 5.5 fits in your automation stack, or want to map out which workflows are now economically viable to automate, book a 15-minute consultation — we'll work through the numbers with you.