Claude 3.5 Haiku

Claude 3.5 Haiku is the small, fast tier of Anthropic's Claude 3.5 model family, announced October 22, 2024 alongside an upgraded Claude 3.5 Sonnet and the public-beta "computer use" capability, and made generally available in early November 2024 (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). Anthropic positioned it as the next generation of its fastest model, stating that at a speed similar to the previous Claude 3 Haiku it improved across every skill set and surpassed Claude 3 Opus, the largest model of the prior generation, on many benchmarks (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). It competed in the low-latency, low-cost tier against GPT-4o mini and Google's Gemini 1.5 Flash. It is the model that Anthropic traced at circuit level in On the Biology of a Large Language Model (March 2025), a public mechanistic-interpretability study of a production frontier model. (Source: On the Biology of a Large Language Model, Mechanistic Interpretability)

Lineage and history

Claude 3.5 Haiku was the third model released in Anthropic's Claude 3.5 generation. The family opened with the original Claude 3.5 Sonnet in June 2024, followed by an upgraded Claude 3.5 Sonnet (sometimes referred to as Claude 3.5 Sonnet "new") announced October 22, 2024, with which Claude 3.5 Haiku was announced (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). A Claude 3.5 Opus had been expected to complete the three-tier structure used in the earlier Claude 3 generation, but Anthropic did not release a 3.5 Opus; later Anthropic statements indicated the planned Opus-tier 3.5 model did not ship under that name (Source: https://blog.promptlayer.com/everything-we-know-claude-3-5-opus/). The three-tier naming convention (Haiku for the small/fast model, Sonnet for the mid tier, Opus for the largest) carried forward from Claude 3.

Within the small-model line, Claude 3.5 Haiku succeeded Claude 3 Haiku (claude-3-haiku-20240307, generally available March 2024) and was itself succeeded by Anthropic's later Haiku tiers. When Anthropic deprecated Claude 3.5 Haiku on December 19, 2025, it named Claude Haiku 4.5 (claude-haiku-4-5-20251001) as the recommended replacement (Source: https://platform.claude.com/docs/en/about-claude/model-deprecations). Claude 3.5 Haiku was retired from Anthropic-operated platforms on February 19, 2026, after which API requests to the model fail; Anthropic-operated retirement followed at least 60 days after the deprecation notice, and partner platforms such as Amazon Bedrock and Google Cloud set their own schedules (Source: https://platform.claude.com/docs/en/about-claude/model-deprecations).

Capabilities and benchmarks

At launch Anthropic reported that Claude 3.5 Haiku exceeded the original Claude 3 Opus on most benchmarks despite being a much smaller model, attributing the gain to compounding post-training and data improvements over the roughly six months between the two models. On coding (HumanEval) it matched or exceeded Opus. Anthropic highlighted coding in particular, reporting that Claude 3.5 Haiku scored 40.6% on SWE-bench Verified, which it said outperformed many agents using publicly available state-of-the-art models, including the original Claude 3.5 Sonnet and GPT-4o (Source: https://www.anthropic.com/news/3-5-models-and-computer-use).

Anthropic described the model as well suited, given its low latency, improved instruction following, and more accurate tool use, to user-facing products, specialized sub-agent tasks, and generating personalized experiences from large volumes of data such as purchase history, pricing, or inventory records (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). Independent evaluation was more mixed: some users reported that on the LiveBench aggregate Claude 3.5 Haiku scored below both Claude 3 Opus and Google's Gemini 1.5 Flash, while costing substantially more per token than Flash (Source: https://www.reddit.com/r/ClaudeAI/comments/1gkgvjb/claude_35_haiku_performs_worse_than_claude_3_opus/).

Training and architecture

Anthropic disclosed little about Haiku's architecture. From public material it is a dense decoder-only Transformer, smaller than Claude 3 Haiku (which was already small) yet more capable at similar or smaller size, reflecting efficiency gains from improved data and post-training. It was trained via Constitutional AI and reinforcement learning from AI feedback (RLAIF), Anthropic's standard alignment stack. The parameter count is undisclosed; third-party estimates range from 20B to 70B. The model carried the 200k-token context window common to the Claude 3 and 3.5 lines. It launched as a text-only model, with Anthropic stating image input would follow (Source: https://www.anthropic.com/news/3-5-models-and-computer-use).

Availability and pricing

Claude 3.5 Haiku was made available across Anthropic's first-party API, Amazon Bedrock, and Google Cloud's Vertex AI, initially as a text-only model, and was also exposed through the Claude app (Source: https://www.anthropic.com/news/3-5-models-and-computer-use).

Anthropic set Claude 3.5 Haiku's launch pricing higher than that of Claude 3 Haiku. The prior Claude 3 Haiku had been priced at $0.25 per million input tokens and $1.25 per million output tokens; Claude 3.5 Haiku launched at $1 per million input tokens and $5 per million output tokens, four times the previous generation's rate (Source: https://www.reddit.com/r/ClaudeAI/comments/1gjl6yl/haiku_35_released/) (Source: https://www.linkedin.com/posts/richardskinner-phasedai_anthropic-has-raised-the-price-of-claude-activity-7259593452156334082-0mfh). The increase drew criticism from developers who had expected the small-tier model to remain inexpensive, given that Claude 3.5 Haiku scored below cheaper competitors such as Gemini 1.5 Flash on some independent benchmarks (Source: https://www.reddit.com/r/ClaudeAI/comments/1gkgvjb/claude_35_haiku_performs_worse_than_claude_3_opus/).

On December 3–5, 2024, Anthropic lowered the price to $0.80 per million input tokens and $4 per million output tokens across all platforms, a roughly 20% reduction announced within a post about its Trainium2 partnership with AWS (Source: https://simonwillison.net/2024/Dec/5/claude-35-haiku-price-drops-by-20/) (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). The same announcement introduced a faster variant of Claude 3.5 Haiku, powered by AWS Trainium2 and offered through Amazon Bedrock cross-region inference in the US East (Ohio) region, with up to 60% faster inference at the original $1 / $5 pricing (Source: https://simonwillison.net/2024/Dec/5/claude-35-haiku-price-drops-by-20/).

Safety and evaluations

The Claude 3.5 models were evaluated under Anthropic's Responsible Scaling Policy. For the upgraded Claude 3.5 Sonnet announced in the same release, Anthropic stated that it evaluated the model for catastrophic risks and found the ASL-2 Standard remained appropriate, and that joint pre-deployment testing was conducted with the US AI Safety Institute and the UK AI Safety Institute (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). Anthropic's published model-card addendum for the October 2024 models documented the Claude 3.5 family's evaluations (Source: https://www.anthropic.com/news/3-5-models-and-computer-use).

Role in interpretability research

Anthropic selected Haiku rather than Opus or Sonnet for its mechanistic-interpretability work for practical reasons: Haiku is small enough that the circuit-tracing, attribution-graph, and feature-extraction techniques developed by Anthropic's interpretability team were computationally tractable at scale. The resulting analysis in On the Biology of a Large Language Model (Anthropic, March 2025) identified:

• Multi-step reasoning circuits, including a case of Haiku computing via an intermediate concept it did not verbalize.
• Planning circuits in poetry generation, in which Haiku chose rhyming words before producing the line that rhymed.
• Mental arithmetic via look-up circuits operating in parallel with symbolic-reasoning circuits.
• "Refusal" features and their interaction with jailbreak prompts.
• Hallucination circuits, in which a "don't-know" feature was suppressed by competing features.

The study was the first time a full production frontier model was traced at circuit level, and Anthropic presented it as establishing Mechanistic Interpretability as an empirically grounded research program rather than a theoretical one. (Source: On the Biology of a Large Language Model)

The work introduced attribution graphs as a tool for tracing information flow through Haiku's residual stream, and developed and demonstrated cross-layer transcoders on Haiku. The paper includes a public catalog of model features, which Anthropic describes as the largest for any frontier LLM. Several circuits identified in Haiku were subsequently confirmed to exist in larger Claude models, which Anthropic cites as evidence that interpretability findings on small models generalize upward. The interpretability work also includes tracing how Constitutional-AI-installed values are represented at the circuit level. (Source: On the Biology of a Large Language Model, Mechanistic Interpretability)

Reception

Anthropic's claim that the model matched or surpassed Claude 3 Opus at far smaller size drew attention as an illustration of how data and post-training improvements can advance a small model's quality within a model generation (Source: https://www.anthropic.com/news/3-5-models-and-computer-use). The launch pricing was the most contested aspect of the release: developers objected to the fourfold increase over Claude 3 Haiku, and some noted that on independent aggregates such as LiveBench the model trailed cheaper alternatives (Source: https://www.reddit.com/r/ClaudeAI/comments/1gjl6yl/haiku_35_released/) (Source: https://www.reddit.com/r/ClaudeAI/comments/1gkgvjb/claude_35_haiku_performs_worse_than_claude_3_opus/). The subsequent 20% price reduction in December 2024 partially addressed the cost objection while introducing a separately priced faster Trainium2 variant (Source: https://simonwillison.net/2024/Dec/5/claude-35-haiku-price-drops-by-20/). The model later gained durable significance less for its commercial performance than for its role as the subject of Anthropic's circuit-level interpretability study. (Source: On the Biology of a Large Language Model)

Related models

Claude 3.5 Haiku sits between Claude 3 Haiku (its predecessor in the small-model line) and Anthropic's later Haiku tiers, with Claude Haiku 4.5 named as its recommended replacement at deprecation (Source: https://platform.claude.com/docs/en/about-claude/model-deprecations). Within its own generation it shared the Claude 3.5 family with the original and upgraded Claude 3.5 Sonnet. Anthropic benchmarked it against the prior-generation flagship Claude 3 Opus and positioned it in the small-model tier against GPT-4o mini and Gemini 1.5 Flash. Circuits first identified in Claude 3.5 Haiku were later confirmed in larger Claude models, a line of interpretability work that continued into models such as Claude Opus 4.6. (Source: On the Biology of a Large Language Model)

Relationships

• instance-of: General-Purpose AI (GPAI)
• subject-of: On the Biology of a Large Language Model
• depends-on: Constitutional AI
• related: Anthropic, Mechanistic Interpretability, Claude 3 Opus, Claude Opus 4.6, GPT-4 Family (OpenAI)