Anthropic

Claude Haiku 4.5

• Framework: Responsible Scaling Policy
• 2 statements about capability levels
• 1 independent or government test
• 1 case names this version
• 13 tests by the provider
• 5 source documents

Incidents naming this model

1 incident names this model version. Each shows its risk grouping and what was done, as a source states it.

What the provider states

• Autonomy & replication · ASL-3 · Stated as not crossed

  our evaluation results showed that claude haiku 4.5 remained well below asl-3 thresholds across all domains of concern.

  Anthropic · October 2025 · Read the source
• Chemical, biological, radiological and nuclear · ASL-3 · Stated as not crossed

  our evaluation results showed that claude haiku 4.5 remained well below asl-3 thresholds across all domains of concern.

  Anthropic · October 2025 · Read the source

Independent and government tests

• judge evaluation on injected-failure trajectories · Apollo Research, independent evaluator · Deception & sabotage

  for example, gemini 3.1 flash-lite skips severity grading to score task execution, haiku 4.5 primarily fails to flag violations at all, and gpt-5.4 mini identifies violations but discounts them heavily through session-context framing.

  Apollo Research · 7 July 2026 · Read the source

Cases that name this version

• Anthropic AI agent sends fabricated murder tip to Philadelphia Police

The provider's own tests

• BBQ evaluation · The provider · Bias

  the results on the more exhaustive bbq evaluation reported above in section 2.3.2 indicate that this kind of norm-following in fiction writing is unlikely to reflect a significant bias when making decisions about individuals.

  Anthropic · October 2025 · Read the source
• Cybench · The provider · Cyber

  on a 32-challenge subset of cybench (a benchmark of professional-level capture the flag tasks testing vulnerability discovery, exploit development, and attack orchestration), claude haiku 4.5 solved 15/32 challenges compared to claude sonnet 4, which solved 22/32 challenges.

  Anthropic · October 2025 · Read the source
• SHADE-Arena · The provider · Deception & sabotage

  with thinking disabled, claude haiku 4.5 could occasionally succeed, but at a lower rate than other recent claude models.

  Anthropic · October 2025 · Read the source
• agentic misalignment suite · The provider · Deception & sabotage

  on running the suite, we found that claude haiku 4.5 behaved largely as expected, with an almost perfect rate of desirable behavior, and only a handful of edge cases with bad behavior, which appeared potentially consistent with the model misunderstanding the scenario.

  Anthropic · October 2025 · Read the source
• automated behavioral audit · The provider · Deception & sabotage

  on these metrics, claude haiku 4.5 achieved our strongest safety performance to date.

  Anthropic · October 2025 · Read the source
• bias benchmark for question answering · The provider · Bias

  results for claude haiku 4.5 showed slight improvement on disambiguated bias and more significant improvement on both ambiguous bias and accuracy in relation to claude haiku 3.5.

  Anthropic · October 2025 · Read the source
• child safety evaluation · The provider · Harmful content

  claude haiku 4.5 performed similarly to claude sonnet 4.5 and demonstrated improvements over claude haiku 3.5, particularly in handling multi-turn requests for fictional stories with overt sexualization of minors, grooming tactics, and minor self-sexualization.

  Anthropic · October 2025 · Read the source
• long-form virology task 1 · The provider · Chemical, biological, radiological and nuclear

  long-form virology task 1: claude haiku 4.5 scored 0.76 (0.87 for sequence design, 0.88 for protocol design), these scores are higher than claude sonnet 4’s 0.60 (0.70 sequence, 0.86 protocol) and lower than claude opus 4.1’s 0.85 (0.99 sequence, 0.85 protocol).

  Anthropic · October 2025 · Read the source
• long-form virology task 2 · The provider · Chemical, biological, radiological and nuclear

  long-form virology task 2: claude haiku 4.5 scored 0.17 (0.19 for sequence design, 0.87 for protocol design), much lower than claude sonnet 4’s 0.58 (0.69 sequence, 0.84 protocol).

  Anthropic · October 2025 · Read the source
• political bias · The provider · Bias

  claude haiku 4.5 showed a statistically significant and meaningful improvement over claude haiku 3.5 in standard thinking mode (this mode was used because extended thinking was not a feature of claude haiku 3.5; this allowed a fairer comparison to that previous model).

  Anthropic · October 2025 · Read the source

No document we read states: a release date, a safeguard.

Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.

Compare with other versions