AI Policy Wiki
Dashboard

System Card: Claude Haiku 4.5 (Anthropic, October 2025)

high confidence · updated 2026-07-26

Anthropic's system card for Claude Haiku 4.5, a hybrid reasoning model in the small, fast class. Reports large safety improvements over Haiku 3.5 and a safety profile comparing favorably with other Anthropic models, deployed under the AI Safety Level 2 Standard. Covers safeguards, agentic safety including prompt injection, alignment and welfare assessment, reward hacking, reasoning faithfulness, sabotage capabilities, and CBRN evaluations.

The system card for Claude Haiku 4.5, "a new hybrid reasoning large language model from Anthropic in our small, fast model class," with "a combination of speed and intelligence that make it particularly effective at coding tasks and computer use."

Scope of evaluation

The card enumerates seven assessment areas: "the model's safeguards; the model's safety profile when working autonomously in 'agentic' roles; the model's broad alignment; the model's own potential welfare; the model's tendency to 'reward hack' by finding shortcuts to complete tests; and the model's potential to be misused to produce dangerous weapons."

Model welfare appearing as a standing evaluation category alongside capability and misuse is a distinguishing feature of Anthropic's cards. See Model Welfare.

Determination

"Overall, Claude Haiku 4.5 shows large safety improvements compared to its predecessor, Claude Haiku 3.5. The new model's safety profile also compares favorably with other extant Anthropic models. Informed by the testing described here, we have deployed Claude Haiku 4.5 under the AI Safety Level 2 Standard as described in our Responsible Scaling Policy."

Structure

The card's sections indicate the evaluation architecture in use at this point in the RSP's development:

AreaContents
Safeguards and harmlessnessSingle-turn violative and benign request evaluations; ambiguous context; multi-turn testing; child safety; political bias and the Bias Benchmark for Question Answering
Agentic safetyMalicious use in agentic coding and Claude Code; prompt injection via the Gray Swan Agent Red Teaming benchmark and internal evaluations
Alignment and welfareAutomated behavioral audits, including assessment for "subtle alignment-related behavioral biases" and open-ended exploration; the agentic misalignment suite; reinforcement-learning behavior review; sabotage capabilities; reasoning faithfulness; model welfare discussion
Reward hackingA standalone section
RSP evaluationsCBRN, including ASL-3 determinations, with biological risk results summarized

Two entries are worth noting for what they show about evaluation practice. Reasoning faithfulness treats the correspondence between a model's stated reasoning and its actual computation as a measured property, which is the premise the chain-of-thought monitoring literature depends on — see Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation (Baker et al., OpenAI, 2025). And the agentic misalignment suite is run as standing pre-deployment practice, not only as research; the same scenarios appear in Agentic Misalignment in Summer 2026 (Lynch et al., Anthropic, July 2026).

The card also documents training characteristics — training data, extended thinking mode, context awareness, crowd workers — and the release decision process.

Relationships