Model
Model explorer

Claude Haiku 5

CLOSED
Anthropic · Claude 5 family · released Sep 17, 2026

The volume tier of the Claude 5 family, released September 17, 2026 — six weeks after Fable 5.1/Mythos 5.1 and the first Haiku at the 5-series capability level. $1/$5 per Mtok makes it a fifth of Sonnet 5's price for roughly 85-90% of its benchmark profile, which Anthropic itself frames as the high-volume-agent play. The launch table leans on the established suite (no DeepSWE or OSWorld 2.0 rows) and notably does NOT repeat the 'best in class' framing Anthropic used for Opus 5 — the numbers are honest for a small tier. 500K context (vs. the family's usual 1M) is the clearest architectural tell.

ReasoningCodingVisionFunction callingTool useAgentic
2153.6
Elo · rank #73
Parameters
Undisclosed
Active params
—
Context
500K tokens
Architecture
Proprietary smallest-tier model of the Claude 5 family; 500K-token context, extended thinking supported
License
Proprietary
Languages
—
API price (in/out)
$1 / $5
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-AGI-2Reasoning61.5%#16
best: GPT-6 Astra · 94.6%
BrowseCompAgents76.3%#29
best: Kimi K3.1 · 92.4%
CharXivVision71.2%#33
best: Qwen3.8-Flash-Next · 90.6%
GDPval-AAAgents1389#36
best: Claude Fable 5 · 1932
GPQA DiamondReasoning86.9%#62
best: GPT-6 Astra · 96.0%
Humanity's Last ExamReasoning39.8%#38
best: Claude Opus 5 · 64.7%
MMLU-ProKnowledge83.4%#39
best: Claude Fable 5 · 91.5%
OSWorld-VerifiedAgents71.6%#20
best: Claude Fable 5 · 85.0%
SWE-bench ProCoding52.8%#52
best: Claude Fable 5.1 · 81.2%
SWE-bench VerifiedCoding78.4%#32
best: Claude Opus 5 · 96.0%
Terminal-Bench 2.0Coding64.7%#45
best: Gemini 3.8 Flash · 89.4%
API price $1/$5 · each benchmark row carries its own source badge (see methodology)