AnthropicReasoning model

Claude Haiku 4.5

Name: Claude Haiku 4.5 benchmark results
Creator: BenchLeader

Best configuration ranks #336 of 610 on the BenchLeader Index at 48.0 ±4.5 (thinking reasoning effort). Last measured 5 Sept 2026. Released 15 Oct 2025.

Blended price: $2.00/M; $1.00 in · $5.00 out
Output speed: 83 tok/s
First answer: 22 s; first token 0.77 s
Context: 200k

How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.

Overall index48
Reasoning39
Coding30
Agents & tools45
Maths58
Knowledge52
Instruction following48
Multimodal43
Long context64
Composite51

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Default” means the publisher did not say which setting was used.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
high	–	–	79 tok/s	0.77 s	$0.0019	Coding 59
thinkingbest	48.0	#336	83 tok/s	22 s	$0.0019	Agents & tools 45 · Coding 30 · Composite 51 · Instruction following 48 · Knowledge 52 · Long context 64 · Maths 58 · Multimodal 43 · Reasoning 39
default	47.9	#341	79 tok/s	0.77 s	$0.0019	Agents & tools 45 · Coding 49 · Composite 48 · Human preference 50 · Instruction following 44 · Knowledge 46 · Long context 51 · Maths 61 · Multimodal 39 · Reasoning 44

Benchmark results

One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Reasoning

Benchmark	high	thinking	default	Source
GPQA Diamond	–	–	71.2%#154	Epoch AI Benchmarking Hub
LMArena Hard Prompts	–	–	1441#105	LMArena
GPQA Diamond (AA)not in index	–	67.2%#307	64.7%#329	Artificial Analysis
Humanity's Last Exam (AA)not in index	–	10.4%#251	4.2%#441	Artificial Analysis
GPQA Diamond (Vals)not in index	–	72.2%#85	–	Vals AI
ARC-AGI-1	–	47.7%#125	14.3%#163	ARC Prize
ARC-AGI-2	–	4.0%#127	1.3%#153	ARC Prize

Coding

Benchmark	high	thinking	default	Source
SciCode	–	–	43.3%#92	SciCode
WeirdML	–	–	45.4%#82	WeirdML
LMArena Coding	–	–	1481#90	LMArena
LMArena WebDev	–	–	1329#97	LMArena
SciCode (AA)not in index	–	42.3%#105	–	Artificial Analysis
LiveCodeBench	–	41.2%#124	–	Vals AI
IOI	–	6.2%#42	–	Vals AI
SWE-bench (Vals)not in index	–	66.6%#68	–	Vals AI
SWE-Bench Pro	–	–	39.5%#12	Scale AI SEAL
SWE-bench Verified (bash only)	66.6%#16	–	–	SWE-bench
SWE-bench Verified (any scaffold)not in index	66.6%#20	–	–	SWE-bench

Agents & tools

Benchmark	high	thinking	default	Source
Terminal-Bench	–	–	35.5%#41	Terminal-Bench
APEX-Agents	–	–	8.9%#53	Mercor
Terminal-Bench Hard	–	27.3%#123	27.3%#123	Artificial Analysis
τ²-Bench Telecom (AA)not in index	–	54.7%#180	32.5%#241	Artificial Analysis
Terminal-Bench 2.1 (Vals)	–	43.8%#53	–	Vals AI
MCP Atlas	–	–	40.2%#30	Scale AI SEAL
BFCL Overall	–	–	68.7%#6	Berkeley Function Calling Leaderboard

Maths

Benchmark	high	thinking	default	Source
OTIS Mock AIME	–	–	66.7%#134	Epoch AI Benchmarking Hub
MATH Level 5	–	–	96.4%#11	Epoch AI Benchmarking Hub
AIME (Vals)	–	82.7%#44	–	Vals AI
MGSM	–	92.2%#21	–	Vals AI

Knowledge

Benchmark	high	thinking	default	Source
SimpleQA Verified	–	–	13.2%#68	Epoch AI Benchmarking Hub
AA-Omniscience	–	-4.4#114	-7.6#126	Artificial Analysis
MMLU-Pro	–	78.7%#96	–	Vals AI
LegalBench	–	81.2%#75	–	Vals AI
CorpFin	–	60.6%#66	60.3%#69	Vals AI
TaxEval	–	67.5%#103	–	Vals AI
MedQA	–	79.6%#71	–	Vals AI

Instruction following

Benchmark	high	thinking	default	Source	Trend
IFBench	–	54.3%#151	42.0%#234	Artificial Analysis
MultiChallenge	–	50.5%#23	–	Scale AI SEAL

Human preference

Benchmark	high	thinking	default	Source	Trend
LMArena Text	–	–	1413#117	LMArena
EQ-Bench 4	–	–	1064#25	EQ-Bench

Multimodal

Benchmark	high	thinking	default	Source	Trend
MMMU-Pro	–	58.5%#183	55.1%#194	Artificial Analysis

Long context

Benchmark	high	thinking	default	Source	Trend
AA-LCR	–	74.3%#116	49.7%#256	Artificial Analysis

Composite

Benchmark	high	thinking	default	Source
Epoch Capabilities Indexnot in index	–	–	142.4#88	Epoch AI Benchmarking Hub
AA Intelligence Index	–	17.6#198	15.4#216	Artificial Analysis
Vals Indexnot in index	–	22.9#47	–	Vals AI

Where to run it

Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.

Provider	Speed	First token	Input $/M	Output $/M	Context	Quantisation
Amazon Bedrock (EU)	97 tok/s	0.99 s	$1.10	$5.50	200k	–
Google Vertex (Europe)	74 tok/s	0.47 s	$1.10	$5.50	200k	–
Amazon Bedrock (US)	72 tok/s	1.36 s	$1.10	$5.50	200k	–
Google Vertex	67 tok/s	0.71 s	$1.00	$5.00	200k	–
Amazon Bedrock (Global)	65 tok/s	0.71 s	$1.00	$5.00	200k	–
Anthropic	59 tok/s	0.70 s	$1.00	$5.00	200k	–
Azure	53 tok/s	0.71 s	$1.00	$5.00	200k	–

Price history

Listed price per 1M tokens over time, as recorded by OpenRouter for the provider with the longest history.

input outputnow $1.10 in · $5.50 out

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.100 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0019	$0.0016	25.8 s
Summarise a 30-page report	12,000 / 600	$0.015	$0.0069	29.4 s
Code edit	6,000 / 1,500	$0.013	$0.0095	40.2 s
Agentic coding session	60,000 / 4,000	$0.080	$0.040	1.2 min
Structured extraction	2,000 / 200	$0.0030	$0.0017	24.6 s