Claude 3.5 Haiku
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,300 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,300 words
Add missing citations, update stale details, or suggest a clearer explanation.
Claude 3.5 Haiku is a large language model developed by Anthropic and released in November 2024 as the fastest and most affordable member of the Claude 3.5 generation. It was positioned as a successor to the original Claude 3 Haiku, built for low-latency, high-volume tasks such as user-facing assistants, sub-agent routines, and large-scale data processing, while reaching benchmark scores that in several cases exceeded Anthropic's earlier flagship model, Claude 3 Opus.[1][2][3]
Anthropic announced Claude 3.5 Haiku on October 22, 2024, in the same blog post that introduced an upgraded version of Claude 3.5 Sonnet and a public beta of the "computer use" capability. At announcement, the company said the model would arrive "later this month" across its first-party API, Amazon Bedrock, and Google Cloud's Vertex AI, and that it would launch as a text-only model with image input to follow.[1][3] The model became available through the Claude API on November 4, 2024, and reached Amazon Bedrock the same day.[2][4][5]
The Claude naming scheme uses three size tiers: Haiku for the smallest and fastest models, Sonnet for the mid-range, and Opus for the largest. Claude 3.5 Haiku occupied the entry tier of the 3.5 family. Its API model identifier is claude-3-5-haiku-20241022, and Anthropic gave it a training-data cutoff of July 2024.[2][6]
Anthropic described Claude 3.5 Haiku as improving "across every skill set" relative to the previous Haiku while keeping similar speed, and highlighted gains in coding, instruction following, and tool use. The company suggested it for user-facing products, specialized sub-agent tasks within larger agentic systems, and generating personalized output from large volumes of data.[1][3]
The model accepts a context window of up to 200,000 tokens and can generate up to 8,192 tokens in a single response.[6][7] At launch it processed text only. Unlike the original Claude 3 Haiku, which accepted images, the 3.5 version shipped without vision, and Anthropic advised users who needed image analysis or the lowest possible cost to continue using Claude 3 Haiku.[2][8] Vision support was added later: Anthropic's API release notes recorded image input arriving for Claude 3.5 Haiku on February 24, 2025, roughly four months after the text-only debut.[9]
Anthropic's central claim was that, despite being the smallest model in the 3.5 generation, Claude 3.5 Haiku surpassed Claude 3 Opus on many intelligence benchmarks at a fraction of the cost and at the speed of the prior Haiku.[1][3] The most cited single figure was on coding: the company reported a score of 40.6% on SWE-bench Verified, a benchmark of real-world software-engineering tasks, and said this placed it ahead of several larger publicly available models at the time, including the original Claude 3.5 Sonnet and OpenAI's GPT-4o.[1][3][8]
Additional benchmark results reported for the model are summarized below. Scores are drawn from Anthropic's materials and aggregators that cite them; exact evaluation settings (shot count, chain-of-thought) vary by benchmark.
| Benchmark | Domain | Claude 3.5 Haiku |
|---|---|---|
| SWE-bench Verified | Software engineering | 40.6% |
| HumanEval | Code generation | 88.1% |
| MMLU-Pro | General knowledge and reasoning | 65.0% |
| MATH | Mathematical problem solving | 69.4% |
| MGSM | Multilingual grade-school math | 85.6% |
| DROP | Reading comprehension | 83.1% |
Sources: Anthropic announcement and model-card addendum for SWE-bench Verified; benchmark aggregators citing the model card for the remaining figures.[1][3][10]
Pricing was the most discussed aspect of the release and went through three distinct stages. When Anthropic first previewed Claude 3.5 Haiku, it indicated that the model would cost the same as Claude 3 Haiku, which was priced at $0.25 per million input tokens and $1.25 per million output tokens.[8][11] At the actual API launch on November 4, 2024, the company instead set the price at $1.00 per million input tokens and $5.00 per million output tokens, a fourfold increase over Claude 3 Haiku on both figures.[2][8][11]
Anthropic justified the higher price by pointing to capability rather than parity with the older model. In a statement posted to its social accounts, the company wrote: "During final testing, Haiku surpassed Claude 3 Opus, our previous flagship model, on many benchmarks, at a fraction of the cost. As a result, we've increased pricing for Claude 3.5 Haiku to reflect its increase in intelligence."[8][12] About a month later, on December 5, 2024, Anthropic reduced the price by 20% to $0.80 per million input tokens and $4.00 per million output tokens, saying the cut was meant to make the model "even more accessible for a wide range of use cases." That lower rate became the model's standard pricing.[13][14]
| Date | Input (per 1M tokens) | Output (per 1M tokens) | Note |
|---|---|---|---|
| Pre-launch indication | $0.25 | $1.25 | Suggested parity with Claude 3 Haiku |
| November 4, 2024 (launch) | $1.00 | $5.00 | 4x increase over Claude 3 Haiku |
| December 5, 2024 | $0.80 | $4.00 | 20% reduction, standard rate thereafter |
The model was distributed through Anthropic's API, Amazon Bedrock, and Google Cloud's Vertex AI.[1][2] Beyond pay-as-you-go API access, it was also offered as the fast option inside Anthropic's commercial products and team plans aimed at organizations, marketed under the Claude for Work banner.
Coverage of the launch concentrated on the pricing change rather than on the model's quality. TechCrunch noted that it is unusual for an AI vendor to raise the price of a model within the same series and that the move raised questions about Anthropic's pricing strategy as competition with OpenAI and Google intensified.[8] Several outlets framed the model as "4x more expensive than its predecessor," and some commentators argued that, because the speed was broadly similar to Claude 3 Haiku, the higher cost was hard to justify for budget-sensitive workloads, particularly given the absence of image support at launch.[2][11][15]
Reaction to the model itself was mixed. The strong SWE-bench Verified result drew attention from developers, and the model found use as a cheaper alternative to mid-tier models for coding and agent tasks where Opus-level reasoning was not required. The subsequent 20% price cut in December was widely read as a response to the earlier criticism, although Anthropic framed it as an accessibility measure rather than a reversal.[13][14]
Claude 3.5 Haiku was later succeeded in the Haiku tier by faster, more capable models in subsequent Claude generations, including Claude Haiku 4.5, which posted substantially higher coding scores while continuing the same emphasis on speed and low cost.[16]