Claude Haiku 4.5

Claude Haiku 4.5 is the small, fast tier of Anthropic's Claude 4.5 generation, released October 15, 2025 and made available to all users the same day. Anthropic positioned it as bringing recently-frontier capability to a lower-cost, higher-speed tier: it reports that Haiku 4.5 matches the prior-generation Claude Sonnet 4 on coding performance at roughly one-third the cost and more than twice the speed, and surpasses Sonnet 4 on some tasks such as computer use (Source: anthropic.com). It is the recommended replacement for Claude 3.5 Haiku in Anthropic's small-model line.

Lineage and positioning

Haiku 4.5 succeeds Claude 3.5 Haiku as Anthropic's small, low-latency model and shares the Claude 4.5 generation with the larger Claude Sonnet 4.5, which Anthropic released two weeks earlier and described as its frontier model and "best coding model in the world" at the time (Source: anthropic.com). Anthropic frames Haiku 4.5 as a near-frontier option for cost-sensitive and latency-sensitive work rather than a frontier model, and as a drop-in replacement for both Haiku 3.5 and the prior-generation Sonnet 4. It is the model Anthropic later named as the recommended replacement when it deprecated Claude 3.5 Haiku on December 19, 2025 (Source: platform.claude.com).

A use pattern Anthropic emphasizes is multi-agent orchestration: a larger model such as Sonnet 4.5 breaks a complex problem into a multi-step plan and then directs a team of Haiku 4.5 instances to complete subtasks in parallel, using Haiku 4.5's speed and lower cost for high-fan-out work (Source: anthropic.com).

Capabilities and benchmarks

Anthropic reports that Haiku 4.5 reaches performance that would have been at the state of the art on its internal benchmarks roughly six months earlier, while running several times faster than Sonnet 4.5 at a fraction of the cost. On Anthropic's reported coding evaluation it scores 73.3% on SWE-bench Verified (Source: anthropic.com; openrouter.ai).

Anthropic's measurement methodology (trial counts, thinking budgets, prompt addenda) is documented in the announcement's footnotes; the reported SWE-bench figure uses a minor prompt addition instructing the model to use tools heavily and write its own tests first (Source: anthropic.com). Early customer evaluations corroborate the near-frontier framing: Augment reported Haiku 4.5 reaching about 90% of Sonnet 4.5's score on its agentic-coding evaluation, Gamma reported 65% instruction-following accuracy on slide-text generation versus 44% from its prior premium-tier model, and GitHub reported quality comparable to Sonnet 4 at faster speed in GitHub Copilot testing (Source: anthropic.com).

Availability and pricing

Haiku 4.5 is available on the Claude apps and Claude Code, and to developers via the Claude API (claude-haiku-4-5), Amazon Bedrock, and Google Cloud's Vertex AI, where Anthropic positions it as a drop-in replacement for Haiku 3.5 and Sonnet 4 at its most economical price point. Pricing is $1 per million input tokens and $5 per million output tokens, a roughly 25% increase over Claude 3.5 Haiku's late-2024 pricing of $0.80/$4 (Source: anthropic.com; caylent.com). Anthropic describes it as well suited to real-time, low-latency tasks such as chat assistants, customer-service agents, and pair programming.

Safety and evaluations

Anthropic released Haiku 4.5 under the AI Safety Level 2 (ASL-2) standard, a lower tier than the ASL-3 standard applied to Sonnet 4.5 and Opus 4.1, on the basis that the model poses only limited risk in the production of chemical, biological, radiological, and nuclear (CBRN) weapons. Anthropic reports that in its automated alignment assessment Haiku 4.5 showed low rates of concerning behaviors, was substantially more aligned than its predecessor Claude Haiku 3.5, and had a statistically significantly lower overall rate of misaligned behaviors than both Claude Sonnet 4.5 and Claude Opus 4.1, which it described as making Haiku 4.5 its safest model yet by that metric. The full reasoning is set out in the Claude Haiku 4.5 system card (Source: anthropic.com; anthropic.com).

Haiku 4.5's position at the bottom of the Claude range gave it an unintended role in an API vulnerability disclosed in August 2026. Stealing Reasoning Traces from Proprietary LLM APIs identified it as the weakest Claude model that both accepts injected encrypted reasoning blocks and supports assistant-turn prefilling, and used it as a decoder: fed the encrypted thinking signature of a stronger sibling such as Opus 4.8, it transcribed that model's hidden reasoning verbatim in plaintext, since it lacks the anti-distillation refusal training applied to the frontier models. A single fixed extraction prompt sufficed, where the equivalent attack on OpenAI's GPT-5.6 Luna required per-block prompt templates and best-of-n sampling. The paper estimates decoding 10,000 traces at roughly $720 at standard Haiku 4.5 rates, and draws the general lesson that "an AI ecosystem's security is only as strong as its weakest link." Anthropic acknowledged the disclosure and the authors report the attack no longer worked afterward.

Related models

• Claude 3.5 Haiku — predecessor in the small-model line; Haiku 4.5 is its recommended replacement.
• Claude Sonnet 4.5 — frontier model of the same Claude 4.5 generation; orchestrates Haiku 4.5 sub-agents.
• Claude 3 Opus — earlier flagship that Haiku 4.5 surpasses on most benchmarks at far lower cost.

The system card reports "large safety improvements compared to its predecessor, Claude Haiku 3.5," with a safety profile comparing favorably to other Anthropic models, and deployment under the AI Safety Level 2 Standard. Its evaluation areas span safeguards, agentic safety (including prompt injection via the Gray Swan benchmark), broad alignment, model welfare, reward hacking, reasoning faithfulness, sabotage capabilities, and CBRN.

Relationships

• developer: Anthropic
• supersedes: Claude 3.5 Haiku — recommended replacement at deprecation.
• related: Claude Sonnet 4.5 — same generation; orchestrator in multi-agent use.
• depends-on: Anthropic's Responsible Scaling Policy (Version 3.1) — Responsible Scaling Policy; released under ASL-2.
• instance-of: General-Purpose AI (GPAI)