Guille Ojeda
Whitepaper: The Business Value Realization of AI
Generative AI & LLMOps
Explore what’s new in Claude Haiku 5.5, including what's improved, where Haiku fits alongside the latest Sonnet, Opus, and Fable models, and what has changed in pricing, API behavior, and safeguards.
Claude Haiku 5.5 has arrived, and it brings substantially stronger capabilities to Anthropic's small-model tier, and a sharp decrease in price. Launched on October 7, 2026, it starts at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, with prompts over 100,000 tokens billed at $0.50 and $2.50 per million input and output tokens. Those token rates are 10% / 50% of Haiku 4.5's, and 5% / 25% of Sonnet 5.5's.
The Haiku model family has always been especially interesting for applications that repeatedly classify messages, extract information, summarize documents, or delegate short tasks to agents. Haiku 5.5's improvements over Haiku 4.5 expand what you can consider for that inexpensive tier, while Sonnet 5.5 remains stronger on many demanding evaluations.
The model is available through the Claude API, Amazon Bedrock and Claude Platform on AWS. In this blog, we'll look at what has improved, where Haiku fits alongside the latest Sonnet, Opus, and Fable models, and what has changed in pricing, API behavior, and safeguards.
A Major Upgrade Over Haiku 4.5
Haiku 5.5's improvements cover both technical and business work. Here are the scores over multiple benchmarks, showing a significant improvement over Haiku 4.5.
These are taken directly from the comparison chart in the Haiku 5.5 system card. Haiku 5.5 and Sonnet 5.5 use maximum effort. Haiku 4.5 predates effort levels and uses its thinking configuration setting.
GDPval-AA evaluates professional deliverables across 44 occupations, including documents, slides, and spreadsheets. Models have tools to complete the work, and their outputs receive relative Elo ratings from pairwise comparisons. AA-Briefcase focuses on source-heavy business projects, judging analytical quality, presentation, and task completion. A higher Elo rating means stronger judged performance. SWE-bench Multilingual tests whether agents resolve repository issues across programming languages. Chartography tests interpretation of specialized professional charts.
The gains over Haiku 4.5 are substantial across all four evaluations. Sonnet 5.5 leads on the more demanding tasks, but Haiku 5.5 now offers much stronger capabilities at a fraction of its token prices. That makes it worth considering for work that can be satisfied with a smaller model.
Haiku 5.5 is the first Haiku model with adjustable effort. Instead of assigning a manual thinking-token budget, you choose a level that steers how thoroughly the model works. Adaptive thinking lets it decide when and how much to reason within the request. That gives you a useful control for matching the work to the task: returning a category and building an analysis from several documents don't need the same depth.
The default effort for Haiku 5.5 is medium. At that effort, its GDPval-AA score is 1277, using about one tenth of the output tokens consumed at max. Its score on Briefcase falls to 1372, using less than one quarter of max's output tokens. Here, max delivers better professional-work results at the cost of many more output tokens.
Context and output capacity also increase: Haiku 5.5 supports 1M context tokens and 128K output tokens, up from Haiku 4.5's 200K and 64K. Those limits match Sonnet 5.5's. The newer tokenizer (the same used by Sonnet 5.5 and Opus 5.5) counts the same text as approximately 30% more tokens than Haiku 4.5, so the nominal context increase doesn't translate into five times as much identical text.
Computer use improves too. On the system card's 82-task OSWorld 2.1 offline subset, Haiku 5.5 reaches 72.4% partial credit and 37.1% full-task completion, versus Haiku 4.5's 15.7% and 1.7%. For comparison, Sonnet 5.5 reaches 83.9% and 48.8%.
Where Haiku 5.5 Fits Best
Haiku 5.5 is a good starting point when the requested result is clearly defined and the work repeats often (eg., classification, extraction, routing, summaries, and short tool workflows).
Say you have a support application that receives a message about a missing delivery. You could use one model call to identify the issue category and extract the order reference, which the application then passes to an existing support process. Haiku would be an excellent choice for that call, since the model has specific information to return and a limited job to perform.
On coding agents, Haiku can handle bounded subagent assignments while a larger model works through the broader problem. This would lead to significant savings on the execution of well-defined assignments such as writing code that passes a set of already defined tests, while maintaining a significant level of intelligence for the parts of the problem that are not yet well defined, such as determining the behavior that said tests need to test for.
Larger models still lead on many of the more demanding evaluations. Terminal-Bench 4.0 tests difficult work in a command-line environment. Published results put Haiku at 39.2%, compared with Fable 5.1 at 55.8% and Opus 5.5 at 66.4%, and Sonnet 5.5 at 70.6%. These results support keeping the larger models for complex agentic coding while assigning Haiku work with a tighter scope.
Haiku 5.5's default effort level is medium. Low effort suits simpler, high-volume calls where speed and token efficiency matter most. Haiku is Anthropic's fastest model at standard speeds, with the explicit exception that Opus Fast Mode is faster.
Haiku 5.5 Pricing
Haiku 5.5 is half the price of Haiku 4.5 on prompts above 100,000 tokens. For prompts below 100,000 tokens, Haiku 5.5's price is 1/10 of Haiku 4.5.
These are Anthropic API list prices in USD per million tokens, verified on October 9, 2026. Provider, Region, and negotiated pricing can differ. Check the latest prices on Anthropic's pricing documentation.
The prompt length determines the rate for the whole request. It includes cached input, and it isn't a marginal surcharge applied only to tokens beyond 100,000, the output rate changes too.
For an example uncached request with 100,000 input tokens and 1,000 billed output tokens, the token charge is $0.0105. Add one input token and the higher band applies, making the charge $0.0525005. Long prompts remain cheaper per token than the other models in the table, but the price hike is significant enough to warrant optimizing prompts to stay below 100,000 tokens where possible.
For a smaller example, 10,000 input tokens plus 1,000 billed output tokens cost $0.0015 on Haiku 5.5, versus $0.015 on Haiku 4.5 and $0.030 on Sonnet 5.5. If you had 100,000 identical requests, those token charges would total $150, $1,500, and $3,000 respectively. This example excludes caching, server-tool fees, and provider differences.
Prompt caching lowers the cost for reused context, although storing it initially has a cache-write cost. For asynchronous work, the Batch API costs half the standard input/output rates.
Updating to Haiku 5.5
Going from Sonnet 5.5 to Haiku 5.5 is pretty straightforward, with only a handful of breaking changes. Moving from Haiku 4.5, however, requires significant changes. Always remember that when moving to a different model, you need to re-test your prompts and workflows, and re-run your evaluations.
The first-party model ID for Haiku 5.5 is claude-haiku-5-5. Here are the breaking changes you'll experience if you just update that model ID, and how to fix them:
The breaking changes from Sonnet 5.5 and Haiku 5.5 are easier to resolve. Haiku 5.5 defaults to medium effort rather than Sonnet 5.5's high. Haiku 5.5 accepts thinking: {"type": "disabled"} at low, medium, or high; at xhigh or max it returns 400. Sonnet 5.5 rejects thinking disabled and instead uses the between_tools parameter to suppress up-front thinking while allowing thinking between tool calls, also through high effort; Haiku 5.5 doesn't accept between_tools, so you need to change it when switching models.
Another thing to note is that responses can begin with thinking blocks, which means you need to select text blocks by type rather than reading content[0] as the answer. Moreover, thinking consumes max_tokens alongside visible text, meaning a small limit can stop generation before an answer appears, and thinking text is omitted by default from the response.
Model compatibility is a separate constraint. Haiku 5.5 doesn't read Sonnet 5.5's thinking blocks, the API drops them and proceeds without that earlier reasoning. Sonnet 5.5 can read Haiku 5.5's blocks on the Claude API, but not on Amazon Bedrock.
Haiku also retains forced tool selection, unlike Sonnet 5.5, Opus 5.5, and Fable 5.1. For example, you can force an extraction tool with an enum-valued category field. The response starts with the tool call, without preceding thinking.
Safeguards and Permitted Usage
Haiku 5.5's cyber controls are stronger than Haiku 4.5's, while allowing broader defensive work than Sonnet 5.5's. Haiku-specific guidance allows source-code vulnerability discovery, but general-access penetration testing and higher-risk dual-use techniques are blocked.
Haiku's biology safeguards use the same harmful-misuse classifiers as Sonnet 5.5. This differs from the broader research-biology classifiers deployed for Opus 5.5 and Fable 5.1. Research biology questions remain available subject to those safeguards.
A classifier decline can arrive as HTTP 200 with stop_reason: "refusal". Check the stop reason rather than treating HTTP success as a completed answer. Sonnet 5.5 can use an opt-in server-side fallback beta on the Claude API for eligible cyber and frontier-LLM declines. Haiku has no server-side fallback or fallback credit, so you must implement retries on the client, and they can incur a new cache write.
Remember that classifier behavior and usage permission are separate. A benign request can be flagged, and another model answering doesn't authorize activity prohibited by the Usage Policy. Similarly, Fable's covered-model retention requirements don't transfer to Haiku. Its retention follows the applicable agreement, platform settings, and any verification-program terms.
Conclusion
Haiku 5.5 makes substantially stronger small-model capabilities available at much lower prices than Anthropic offered before. If you're currently using Sonnet 5.5 for a narrow task, Haiku 5.5 is likely to get you lower costs and faster responses. Sonnet still has a place for more demanding work.
Before replacing an existing workload, verify that Haiku 5.5 can correctly complete it, and compare latency and costs on representative inputs. Always run your own evaluations and don't blindly trust public benchmarks to accurately reflect your own type of work. Consider all the breaking changes when going from Haiku 4.5 to Haiku 5.5, and even from Sonnet 5.5 to Haiku 5.5.
How Caylent Can Help
As a Preferred Services Partner and Charter Member of the Claude Partner Network, Caylent helps you select and implement Claude models for your applications. Our Anthropic Consulting & Engineering practice supports adoption, enablement, and building agentic applications, including the workload evaluation and cost optimization needed to make a model upgrade useful in production.
If you're upgrading a Haiku application or considering Haiku 5.5 for work currently handled by Sonnet 5.5, connect with Caylent to assess the model's fit and plan the integration around your application's requirements.
Accelerate your GenAI initiatives
Leveraging our accelerators and technical experience
Browse GenAI Offerings
Related Blog Posts
Claude Sonnet 5.5: Better Coding and Knowledge Work at Sonnet 5 Prices
Explore what’s new in Claude Sonnet 5.5, including its improvements, where it fits alongside other Claude models, and what organizations should know before migrating.
Generative AI & LLMOps
Report: Caylent Named a Major Player in the IDC MarketScape
Generative AI & LLMOps
