AI Potluck
Model components / Fine-tuned / chat models

Claude Haiku

Anthropic

Anthropic's small, fast Claude tier, built for high-volume low-latency coding and chat where cost and speed matter more than peak reasoning. Sits beneath Sonnet and Opus. Tracked as a tier rather than a point release, with the measured release named in the score note. Closed weights, available through the Anthropic API and the Claude apps.

Tier product, currently shipping Haiku 4.5. Renamed from claude-haiku-4-5 so the slug survives the next generation. Verified 2026-08-13 via the Claude Haiku product page.

Openness

1 high confidence
1.0
weights
closed
data
closed
code
closed
license
Proprietary(API-only)

Closed by construction and stable across generations. Measured against Claude Haiku 4.5, the generation shipping at the time of scoring. Haiku 4.5 is sold by the token at $1 per million input tokens and $5 per million output tokens across Claude.ai, the Claude Platform, Bedrock, Vertex AI and Microsoft Foundry, with no weight download offered.

  • https://www.anthropic.com/claude/haiku recorded 2026-08-13

    "Pricing for Haiku 4.5 on the Claude Platform starts at $1 per million input tokens and $5 per million output tokens"; distributed via Claude.ai, the Claude Platform, Amazon Bedrock, Vertex AI and Microsoft Foundry; no weights, corpus or training code offered

Adoption

4 medium confidence
4.0

High-volume tier by design, but no per-model usage figure is published and it is chosen less often than Sonnet as a default. Directional.

  • https://www.anthropic.com/claude/haiku recorded 2026-08-13

    "advanced capabilities at a price that makes Claude practical for scaled deployments, from powering free tier products to budget-conscious applications"; latency-sensitive, coding sub-agent and research sub-agent use cases; no usage or user figure published

Capability

4 medium confidence
4.0

Haiku 4.5 is strong for its size class and competitive with the previous Sonnet generation, but below the current Sonnet, Opus and Fable tiers. Confidence medium - SWE-bench Verified's leaderboard has no submissions since 2025-12-15, so figures are vendor-reported. Anthropic reports 73.3% on SWE-bench Verified and parity with Sonnet 4 on coding, computer use and agent tasks, which is below the current Sonnet generation and well below Opus and Fable - exactly the position the 4 encodes.

Verified 2026-08-13