Skip to content
Leaderboard / Anthropic

Claude Haiku 4.5

Anthropic 2025-10-01 CLOSED Proprietary
Verdict

Claude Haiku 4.5 is behind the leaders: #216 of 406 by SIscope score. Strongest at following instructions, weaker at reasoning & knowledge. Mid-range on price.

SIscope score
44
#216 of 406
Human-vote rank
#162
of 419 · score 1397
Price per 1M tokens, blended
$2
$1 in · $5 out
Context window
200K
200,000 tokens

Shape

Each spoke is a category, 0 at the centre, 100 at the edge.

CodingAgentsno data yetReasoningMathChatInstructions

Category scores

Percentiles: 90 means better than about 90% of tested models.

Coding47
Agents & toolsNo data yet–
Reasoning & knowledge36
Math43
Chat quality62
Following instructions77

What goes into each category · Compare with DeepSeek V4.1 Flash

Human-vote score over time

Text arena, recent updates.

139013981405Oct 20: 1397 (rank #158)Oct 22: 1396.8 (rank #161)Oct 28: 1396.9 (rank #162)Oct 20Oct 28

Coding votes over time

WebDev arena, recent updates.

132513301335Oct 16: 1330.2 (rank #120)Oct 18: 1330.4 (rank #121)Oct 19: 1330.4 (rank #123)Oct 22: 1330.5 (rank #125)Oct 23: 1330.5 (rank #126)Oct 27: 1330.4 (rank #127)Oct 28: 1330.4 (rank #128)Oct 16Oct 28

Independent test runs

Run by Epoch AI on the same harness for every model. Bars show the share of questions answered correctly.

GPQA Diamond60%
PhD-level science multiple choice questions
Better than 33% of 333 tested models. Run Oct 16, 2025.
SimpleQA Verified13%
Short factual questions; rewards not making things up
Better than 8% of 100 tested models. Run Aug 10, 2026.
OTIS Mock AIME 2024-2536%
Competition maths problems in the style of AIME
Better than 27% of 313 tested models. Run Oct 16, 2025.

Human-vote arenas

People compare two anonymous answers and pick the better one. Scores work like chess ratings.

ArenaRankScoreRange (95%)VotesUpdated
Text arena
Everyday chat, blind human votes
#162 of 41913971395–1399154,063Oct 28, 2026
WebDev arena
Building web apps, blind human votes
#128 of 15213301325–133528,961Oct 28, 2026
Vision arena
Questions about images, blind human votes
Not on this board yet
Agent arena
Real multi-step agent sessions
Not on this board yet

Source: Arena (formerly LMArena) leaderboard dataset, CC BY 4.0. Fetched Nov 4, 2026.

Price

Input
$1
per 1M tokens
Output
$5
per 1M tokens
Blended
$2
3 in : 1 out, per 1M
Context window
200K
max tokens per request
Speed
Coming soon
our own tests

List price from the OpenRouter models API (anthropic/claude-haiku-4.5), Nov 4, 2026. Providers can charge differently.