Introducing Claude Haiku 5.5 Anthropic
Introducing Claude Haiku 5.5: the most affordable, quickest, and most succesful small mannequin we’ve ever launched.
Claude Haiku 5.5 is designed for high-volume, cost-sensitive duties. It reliably handles fast and repetitive workloads (like summaries, compactions, database queries, and classification requests). It pairs properly with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. And, because it’s additionally our quickest mannequin to this point, it really works particularly properly for speed-sensitive duties like reside buyer help and browser use.¹
Haiku 5.5 is accessible at a a lot cheaper price than Haiku 4.5. On common, it now prices round 75% much less to run.²
Along with this launch, we’re bettering the worth of our mannequin vary. We’re halving the worth of Claude Sonnet 5.5’s cache reads, which suggests Sonnet 5.5 now runs round 20% cheaper on most agentic work. And we’re introducing a brand new month-to-month API credit score for our Claude Max and Team subscribers, designed to help our customers in constructing new brokers and purposes that run on the Claude Platform.
Performance
Here’s how Claude Haiku 5.5 performs throughout a spread of benchmarks:
| Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5For reference | |
|---|---|---|---|---|
| Knowledge workGDPval-AA v2.1 | 1620 | 735 | 1437 | 1840 |
| Knowledge workAA-Briefcase v1.1 | 1578 | 614 | 1336 | 1824 |
| Computer useOSWorld 2.1 | 72.4%Offline subset | 15.7%Offline subset | 48.9%Offline subset | 83.9%Offline subset |
| Multidisciplinary reasoningHumanity’s Last Exam | 45.9%no instruments | 10.2%no instruments | — | 56.9%no instruments |
| 57.4%with instruments | 18.7%with instruments | — | 64.5%with instruments | |
| Agentic codingTerminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| Agentic codingFrontierCode 1.1 (Main) | 46.4% | — | 42.4% | 52.1%Xhigh |
| Visual reasoningChartography | 46.4%no instruments | 6.4%no instruments | 29.1%no instruments | 61.6%no instruments |
For particulars on how we run our evaluations, see the Haiku 5.5 System Card.
Haiku 5.5 is our first Haiku-class mannequin to return with an adjustable effort setting. This implies that, as with our different fashions, customers can resolve whether or not to optimize for value or intelligence. The charts beneath present how Haiku 5.5 performs on three benchmarks at every effort setting:
Computer use: OSWorldKnowledge work: GDPval-AAMultidisciplinary reasoning: Humanity’s Last Exam
In early testing, our prospects reported outcomes in line with the efficiency and value enhancements proven above. Here’s what they instructed us concerning the new mannequin:
AsanaHubSpotAlphaSenseBoxRogoCognition
Pricing
The desk beneath reveals how Claude Haiku 5.5’s pricing compares to our different fashions. Haiku 5.5 is particularly good worth when used for duties with prompts as much as 100,000 tokens, which make up round 90% of requests to our earlier Haiku mannequin.
| Price per 1 million tokens | Haiku 5.5 prompts as much as / over 100k |
Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| Cache reads | $0.01 / $0.05 | $0.10 | $0.10 |
| Cache writes | $0.125 / $0.625 | $1.25 | $2.50 |
| Input tokens | $0.10 / $0.50 | $1.00 | $2.00 |
| Output tokens | $0.50 / $2.50 | $5.00 | $10.00 |
Safety
Alignment. Claude Haiku 5.5 reveals main enhancements throughout nearly all of our alignment evaluations relative to Haiku 4.5. In explicit, we discovered far fewer cases of misaligned conduct, and a decrease willingness to cooperate with misuse. The mannequin’s system card describes our analysis course of and leads to extra element.
Safeguards. Consistent with its capabilities, Haiku 5.5’s cybersecurity safeguards are extra restrictive than Haiku 4.5’s, however considerably much less restrictive than these we’ve utilized to different latest fashions. In cybersecurity, they enable a wider vary of defensive duties than our safeguards for Sonnet 5.5, however they nonetheless block penetration testing and different methods extra seemingly for use by attackers.
Haiku 5.5’s biology safeguards are the identical as for Sonnet 5, Sonnet 5.5, and Opus 5. They permit analysis biology questions however limit entry to requests that we choose as more likely to trigger hurt. Organizations engaged on wider-ranging biology and cyber actions can apply to our Life Sciences Verification Program and Cyber Verification Program.
Availability
Claude Haiku 5.5 is accessible now on all platforms, together with Amazon Web Services, Google Cloud, and Microsoft Azure. On the Claude Platform, builders can get began with claude-haiku-5-5.
See our migration guide for particulars.
Further updates
Alongside our new pricing for Claude Haiku 5.5, we’re making additional enhancements to the worth of our fashions and merchandise.
First, beginning in the present day, we’re reducing the worth of cache reads on Claude Sonnet 5.5. Cache reads now value 50% much less: $0.10 per million tokens fairly than $0.20. Because cache reads make up a big share of fashions’ token consumption, this reduces the price of Sonnet 5.5 on most agentic duties by round 20%.
For occasion, right here’s what the worth reduce means for Sonnet 5.5’s efficiency relative to value on Terminal-Bench 4.0:
This chart illustrates an necessary distinction between Haiku 5.5 and our bigger fashions. Sonnet 5.5 and Opus 5.5 stay higher decisions for complicated agentic coding duties like these measured by Terminal-Bench 4.0. By distinction, Haiku 5.5 is greatest suited to extra narrowly scoped duties that may in any other case have been cost-prohibitive with earlier variations of Claude—like compaction, summarization, or subagent work.
Second, this week, we’ll roll out a brand new month-to-month API credit score to all Max and Team subscribers to be used on the Claude Platform. Max 5x customers will get $100 in credit per 30 days, Max 20x customers will get $200, and Team subscribers will obtain as much as $500, pooled throughout their customers. These credit are designed to permit our customers to experiment with constructing instruments, apps, and brokers that decision our API. They can be utilized on any of our fashions. For extra data, see our Help Center article.
For builders, we’re additionally updating our Claude Python and TypeScript SDKs so as to add help for laptop use and browser use in beta. Haiku 5.5 is particularly well-suited to those duties, given its mixture of pace, functionality, and value. You can learn extra about this in our Claude Platform docs.
