When to use Haiku 5.5, Sonnet 5.5, and Opus 5.5
The three models complement each other. Anthropic positions Haiku 5.5 for narrow, repetitive work, Sonnet 5.5 for well-scoped everyday tasks, and Opus 5.5 for complex, open-ended work.What changes from Claude Haiku 4.5
Stronger on repetitive work
Haiku 5.5 covers the same kind of task as Haiku 4.5 with a much wider margin. On professional work (GDPval-AA), the score published by Anthropic rises from 735 to 1,620. On computer use in the offline subset of OSWorld 2.1, it rises from 15.7% to 72.4%.Faster, with adjustable effort
It is Anthropic’s fastest model at each model’s standard speed. Opus models in Fast Mode can still be quicker. Thinking is adaptive and on by default, with effort set tomedium. At a lower effort level, the model thinks less and can skip thinking on simple requests.
What early tests reported
- Asana measured more than 30% lower latency on task completion and up to 2.5x faster inference per turn, on bug triage, project setup, and portfolio search.
- HubSpot recorded 92.8% on a CRM task suite (average of three runs). On an audit of stale, ambiguous records, Haiku 5.5 was the fastest, with the highest hit rate and the lowest false-positive rate among the models they tested.
- AlphaSense, at about 8 million weekly calls that answer questions over documents, compared 400 queries: 0.84 versus 0.76 for Haiku 4.5.
- Box saw 11 points above Haiku 4.5 at about half the latency, on reports and summaries that run at scale.
- In Devin Fusion, Haiku 5.5 as a sidekick held a FrontierCode score of 66.2 with Opus 5.5 as the lead, at lower latency.
What the benchmarks show
In the results published by Anthropic, the jump over Haiku 4.5 is large in almost every area. On agentic terminal coding, Sonnet 5.5 remains well ahead.
In Artificial Analysis’s independent evaluation, the low-effort variant scored 29 on the Intelligence Index, ranking 22nd among 182 models in the same class. The median for that price range is 13. In that measurement, the model generated text at 181 tokens per second and cost $0.02 per index task. It was also concise: 32 million output tokens on the index, against a median of 100 million.
On BenchLM, the partial aggregate is 66.35 out of 100, ranking 28th among 888 models. Coverage is still partial (18 benchmarks), so the score is conservative. Haiku 4.5 appears at 41.45 on the same ranking.
The table was published by Anthropic. Artificial Analysis’s Intelligence Index measures the low-effort variant, and the BenchLM aggregate uses partial coverage. Results depend on the harness, the tools, and the effort used. Validate in your own workflow before changing a production setup.
Where it doesn’t lead
- On complex agentic coding, such as Terminal-Bench 4.0, Sonnet 5.5 and Opus 5.5 remain clearly stronger (70.6% versus 39.2% for Haiku 5.5).
- On open-ended professional work, Sonnet 5.5 is still ahead (1,840 versus 1,620 on GDPval-AA).
- Time to first token rises while thinking is active. For simple classification and extraction, a lower effort level tends to answer sooner.
Where it adds the most value
A queue that has to be classified all day
Emails, tickets, and messages keep arriving, and each one needs a topic, an urgency, and a destination. Haiku 5.5 applies the same rules at volume, with low latency. Expected result: every item comes out with a category, a priority, and a route, and ambiguous cases are set aside for a person.Short questions over many documents
Someone asks a specific question about one or a few files, thousands of times. The model finds the passage, answers only what was asked, and points to the source. Expected result: short answers, with the source passage cited, without rewriting the whole document.A subagent next to a larger model
Sonnet 5.5 or Opus 5.5 builds the deliverable. Haiku 5.5 steps in only to fetch a number, summarize an attachment, or compact the context before the next step. Expected result: the larger model spends its context on judgment, and the repeated lookups stay with Haiku 5.5.How to apply it in Tess
1
Pick narrow, repetitive work
Prefer classification, routing, extraction, summaries, and subagents. For open-ended work or long coding tasks, prefer Sonnet 5.5 or Opus 5.5.
2
Lock the response format
Send the categories, the examples, and the exact output shape. Say what to do when a case fits none of the categories.
3
Set tools and approvals
Enable only the tools you need. Require confirmation before sending, publishing, or changing data.
4
Use less effort on simple tasks
Short classification and extraction are usually enough at low effort. Save higher effort for tasks that actually need reasoning.
5
Sample before you scale
Run a small batch, check errors and false positives, and only then raise the volume.
Three playbooks to get started
1. Classification and routing
Input: messages or tickets, the category list, and an example of each.Tools: none, or read-only access to the source queue.
2. Extraction with a citation
Input: one or a few documents and the question.Tools: file reading.
3. Lookup subagent
Input: the lead model’s specific question and the files where the answer should be.Tools: file reading.
Limitations and best practices
- For complex, open-ended work or long coding tasks, Sonnet 5.5 and Opus 5.5 remain the safer choice.
- The new tokenizer counts about 30% more tokens than Haiku 4.5 for the same text. An output limit tuned on the previous model can cut the reply short.
- A 1M-token context does not replace selecting information. Send only relevant content.
- Cybersecurity safeguards are more restrictive than Haiku 4.5’s and block penetration testing and techniques more likely to be used in attacks. Defensive tasks have a wider allowance than on Sonnet 5.5.
- Biology safeguards match those on Sonnet 5, Sonnet 5.5, and Opus 5.
- Review a sample before publishing, sending, or automating decisions.
- Keep human approval for financial, legal, clinical, or high-impact decisions.