TL;DR
- In knowledge work, computer manipulation, and reasoning, Claude Haiku 5.5 is much better than Haiku 4.5 and slightly ahead of Sonnet 5.5 in the published coding benchmarks.
- Computer use: 72.4% for Computer use, 15.7% for Haiku 4.5.
- Reasoning and coding: 45.9% on Humanity's Last Exam (no tools) and 39.2% on Terminal-Bench 4.0.
- According to the provided pricing information, API rates start at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens.
- Best fit: It may be explored for high-volume classification and summarization, computer automation, and focused coding projects, but it needs real-world testing.
Claude Haiku 5.5 shows significant progress over its predecessor, Haiku 4.5, across computer use, reasoning, knowledge work, and agents. Anthropic launched the model on October 7, 2026, and positioned Haiku 5.5 as the fastest and most cost-efficient small model.
You want top performance at low cost; Haiku 5.5 aims to deliver it. Anthropic calls it the most capable Haiku yet; it can handle quick and repetitive workloads involving many parallel sub-agents.
These advantages matter for real applications. Haiku 5.5 benchmarks offer an attractive mix of speed, capability, and cost for high-volume applications.
We review published scores, explain what each benchmark measures, compare Haiku 5.5 to other models, and discuss what the scores mean for developers and businesses.
Claude Haiku 5.5 Benchmarks At A Glance
Anthropic published Haiku 5.5's benchmark results alongside comparisons with GPT-6 Luna, Haiku 4.5 and Sonnet 5.5. The table below summarizes the published figures.
| Benchmark | Category | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|---|
| GDPval-AA v2.1 | Knowledge work | 1,620 | 735 | 1,437 | 1,840 |
| AA-Briefcase v1.1 | Knowledge work | 1,578 | 614 | 1,336 | 1,824 |
| OSWorld 2.1 (offline subset) | Computer use | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity's Last Exam (no tools) | Reasoning | 45.9% | 10.2% | — | 56.9% |
| Humanity's Last Exam (with tools) | Reasoning | 57.4% | 18.7% | — | 64.5% |
| Terminal-Bench 4.0 | Agentic coding | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 (Main) | Agentic coding | 46.4% | — | 42.4% | 52.1% |
| Chartography | Visual reasoning | 46.4% | 6.4% | 29.1% | 61.6% |
Source: Anthropic's official Claude Haiku 5.5 announcement, published October 7, 2026
The table highlights two different strengths. Haiku 5.5 makes substantial progress over its predecessor in knowledge work and computer use, while Sonnet 5.5 retains a clear lead in the more demanding coding and professional-work evaluations.
Where Claude Haiku 5.5 performs best
Haiku 5.5 seems to perform best in computer use and knowledge work. It excels in these areas, making it suitable for processing many requests quickly, but not necessarily for solving a few very hard problems.
1. Knowledge work
Haiku 5.5 scores 1,620 on GDPval-AA v2.1, compared with 735 for Haiku 4.5. The published rating has increased by 885 points. Compared with Sonnet 5.5, which leads the race, GPT-6 Luna scores 1,437, while Sonnet 5.5 scores 1,840.
The second evaluation, AA-Briefcase v1.1, is similar. Haiku 5.5 scores 1,578, ahead of GPT-6 Luna at 1,336 and beats Haiku 4.5 at 614. Sonnet 5.5 remains ahead at 1,824.
GDPval-AA assesses AI performance across 44 business professions. AA-Briefcase evaluates work-oriented tasks. The findings indicate that Haiku 5.5 can handle more complex, professional tasks than Haiku 5.4.
This improvement may benefit reporting, document analysis, information extraction, and recurring tasks. However, a higher benchmark rating doesn’t guarantee accuracy in every workflow. Before production use, test the model against your documents, rules, and expected outputs.
2. Computer use
The offline subset of OSWorld 2.1 rates Haiku 5.5 at 72.4%. Haiku 4.5 scores 15.7%, GPT-6 Luna reaches 48.9%, and Sonnet 5.5 achieves 83.9%.
OSWorld measures an AI agent's ability to run a computer and perform multi-step tasks. The model should not just respond to a question; it must engage with a software environment and act toward a desired outcome.
The changeover to Haiku 4.5 is particularly notable in Anthropic's comparison. It suggests that Haiku 5.5 may be more useful for workflows such as form filling, data entry, navigating interfaces, and moving data between applications.
A model's usefulness isn't determined by its score alone. Real applications must also account for varying interfaces, failed actions, permissions, and errors that can accumulate across steps. Strong benchmark results help, but don’t replace monitoring and recovery processes.
3. Reasoning and use of tools
Haiku 5.5 gets 45.9% on Humanity's Last Exam, without tools and 57.4% with tools. Haiku 4.5 scores 10.2% and 18.7%, respectively, while Sonnet 5.5 scores 56.9% without tools and 64.5% with tools.
Humanity's Last Exam tests higher-order academic achievement and integrated thinking. The tool-assisted results show the model's performance when using external tools to solve the problem.
the difference between the two Haiku 5.5 scores. Tools improve it, allowing it to be used in applications that involve model reasoning and retrieval, calculations, or other structured operations.
However, not all answers are correct simply because they are generated using a tool. Developers must ensure the tools they use produce accurate output, handle missing data, and avoid treating unreliable retrieved content as authoritative.
Claude Haiku 5.5 coding benchmarks: A more mixed result
Haiku performs better than Haiku 4.5 in the published coding evaluations. But it focuses more on task execution than on most agentic workflows.
| Coding benchmark | What it evaluates | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Multi-step command-line tasks | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 (Main) | Coding and software development tasks | 46.4% | — | 42.4% | 52.1% |
What do the coding scores indicate?
Coding scores suggest Haiku 5.5 lags behind larger models on complex terminal tasks. On Terminal-Bench 4.0, which tests agents’ complex terminal skills, it scores 39.2% versus Sonnet 5.5’s 70.6%. These tasks may require extensive planning, debugging, verification, and tool coordination, but Haiku may still suit many software projects.
The gap narrows on FrontierCode 1.1. Haiku scores 46.4%, compared with GPT-6 Luna at 42.4% and Sonnet 5.5 at 52.1%.
Haiku 5.5 suits focused edits, simple bug fixes, code explanations, test generation, and delegated tasks. Larger models may handle architecture or long engineering work better. In multi-agent systems, stronger models can plan and delegate narrow tasks to Haiku, lowering costs.
At this tier the benchmark that matters is cost per million tokens. Our Haiku 5.5 alternatives guide covers the field on that.
Visual reasoning: What Chartography tells us
Haiku 5.5 scores 46.4% on Chartography, versus 6.4% for Haiku 4.5, 29.1% for GPT-6 Luna, and 61.6% for Sonnet 5.5.
Chartography evaluates visual reasoning using charts and related information. These results suggest that Haiku 5.5 can interpret both visual and textual information, supporting tasks such as extracting chart data, understanding visual reports, answering chart-related questions, and processing documents that combine charts and text.
However, it may still misread charts with dense labels, unusual designs, low resolution, or ambiguous scales. For business decisions, verify model-generated interpretations and check extracted figures against the source data.
How Anthropic's benchmark methodology affects the results
Benchmarks compare models, but results depend on test design and execution.
Anthropic's scores correlate with evaluations like OSWorld, Humanity's Last Exam, Terminal-Bench, FrontierCode, GDPval-AA, and Chartography. These tests measure different abilities, so their scores aren’t interchangeable measures of overall intelligence.
Consider four factors when interpreting the results.
- Different scoring systems: GDPval-AA and AA-Briefcase are both based on ratings rather than percentages. You can't directly compare them with accuracy scores on OSWorld or Terminal-Bench.
- Tool availability: Tool availability will produce different outcomes in Humanity's Last Exam. Comparisons need to account for whether models have access to outside information.
- Evaluation settings: Results can be influenced by prompts, agent frameworks, inference effort, and test environments. The headline table does not provide consistent conditions across all models.
- Task coverage: A model might be good at solving a structured computer task but not very good at solving a complex one, handling an unfamiliar interface, or meeting specific domain requirements.
Anthropic provides the headline scores and links to its Haiku 5.5 System Card for more results and evaluation details. Developers should not conclude small score differences.
As of October 8, 2026, the published comparison suggests Haiku 5.5 is a worthwhile upgrade over Haiku 4.5 in some areas. It does not establish Haiku 5.5 as best for every workload or substantiate every claim with reported scores.
Claude Haiku 5.5 pricing and cost per task
One key feature of Haiku 5.5 is its low API price. Anthropic provides various rates, depending on the number of tokens in your prompt: up to 100,000 tokens and beyond.
| API usage | Prompts up to 100K tokens | Prompts over 100K tokens |
|---|---|---|
| Input tokens | $0.10 / million | $0.50 / million |
| Output tokens | $0.50 / million | $2.50 / million |
| Cache reads | $0.01 / million | $0.05 / million |
| Five-minute cache writes | $0.125 / million | $0.625 / million |
*Prices are in Dollars and may vary based on API features and applicable terms.
At the quoted $1/$5 rate, Haiku 5.5 costs 90% less per input/output token than Haiku 4.5 for prompts of up to 100,000 tokens. Differences in tokens processed per second account for the difference in cost estimates, which, based on Anthropic's estimates, is around 75% for Haiku 5.5.
That distinction matters because a lower token price doesn't necessarily translate into the same percentage reduction in total cost. A model may use more tokens, need more attempts, or require a more complex agent workflow to complete a task.
Teams should therefore measure cost per completed task, not cost per million tokens, to assess success. Latency, retry rates, tool calls, and human review can all affect the economics.
For the full rates, read our Haiku 5.5 pricing guide.
Which Claude Haiku 5.5 benchmark matters for your application?
The best indicator will depend on the type of application you are developing. The best model for complex code generation might not be the best model for computer use.
| If you are building | Benchmarks to examine | What the results suggest |
|---|---|---|
| Business reporting and document analysis | GDPval-AA, AA-Briefcase | It can produce great improvement over Haiku 4.5 in knowledge work |
| Browser and desktop automation | OSWorld 2.1 | A substantial gain in computer-use performance |
| Research and tool-assisted assistants | Humanity's Last Exam | Higher scores with tools than without them |
| Coding assistants | Terminal-Bench 4.0, FrontierCode 1.1 | Useful coding capability, but weaker than Sonnet 5.5 on these tests |
| Chart and visual-data analysis | Chartography | Improved visual reasoning over Haiku 4.5 |
| High-volume customer support | Latency, task accuracy, cost per resolution | A promising fit, subject to application-specific testing |
| Multi-agent systems | Coding benchmarks, latency, and cost per task | A potential low-cost option for delegated work |
The pragmatic approach is to test a subset of real tasks to determine the best model for a developer's specific products. Benchmarking may help narrow choices, but production performance should decide.
What can you build with Claude Haiku 5.5?
Haiku 5.5 is fast, cheap and capable, making it a good fit for many fairly specific applications.
- Customer support assistants: Categorize inquiries, gather relevant data, record interactions, and escalate complex inquiries to human agents.
- Document intelligence tools: Summarize reports, answer questions about provided content, compare documents, extract key fields.
- Recurring Reports, CRM Updates, Data Classification, Operational Analysis: Create workflows for internal business applications!
- Browser automation: Automate repetitive browser tasks with proper permissions, logging and error handling.
- AI coding workflows: Automate small coding tasks, create tests, explain code, and support quick coding.
If you are a team developing a product with AI, you can use a model like Haiku 5.5 to assist with specific tasks in a larger application workflow.
Check out our Emergent to learn more about turning product ideas into working apps.
Can you use Claude Haiku 5.5 yet?
Yes. Anthropic's release date for Haiku 5.5 is October 7, 2026, and it will be available on Claude.ai, Claude Code, and in its developer platform.
The model can be called using the Claude API with the model ID claude-haiku-5-5. Availability and configuration may differ from platform to platform.
The model features adaptive thinking with an adjustable effort setting, up to 128,000 output tokens and a maximum context window of 1 million tokens. These are provided to help developers optimize the reasoning depth, speed and cost of the reasoning process in specific workloads.
Benchmark sets don’t always line up across vendors. Our Haiku 5.5 vs Gemini 3.8 Flash comparison covers where they do.
Choose the model that fits your workload
Claude Haiku 5.5 substantially improves on Haiku 4.5 in knowledge work and computer use. Its 72.4% OSWorld score, competitive professional-work rating, and stronger Humanity's Last Exam results make it appealing for fast, cost-effective applications at scale.
The coding results add another important qualification. Haiku 5.5 trails Sonnet 5.5 on both Terminal-Bench 4.0 and FrontierCode 1.1, so it may not be the most efficient option for every complex engineering task.
The most effective approach for builders is to understand each model's capabilities so they can match them to different tasks. If its speed and cost are advantageous, use Haiku 5.5; if it is accurate on real workloads, use it; otherwise, consider routing harder workloads to a larger model if needed.
This isn't just a more affordable AI application. It is a system that allocates model capabilities where they deliver the best value.
Most AI app builders stop at prototypes. Emergent creates production-ready apps you can actually launch.
- Production-ready apps
- Web & mobile apps
- Deploy in minutes
