BenchLeader
Menu
LeaderboardsModelsProvidersBenchmarksCompareReleasesBlogMethodology
xAI

Grok 3

Best configuration ranks #295 of 610 on the BenchLeader Index at 50.0 ±3.9. Last measured 2 Sept 2026. Released 17 Feb 2025.

Blended price
$1.56/M
$1.25 in · $2.50 out
Output speed
–
First answer
–
Context
131k
How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.
  1. Overall index50
  2. Reasoning44
  3. Coding46
  4. Agents & tools34
  5. Maths54
  6. Knowledge52
  7. Instruction following48
  8. Human preference60
  9. Multimodal65
  10. Long context53
  11. Composite44

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.

EffortIndexRankSpeedFirst answerChat reply costCategories
thinking––––$0.0076Composite 42
defaultbest50.0#295––$0.0013Agents & tools 34 · Coding 46 · Composite 44 · Human preference 60 · Instruction following 48 · Knowledge 52 · Long context 53 · Maths 54 · Multimodal 65 · 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

BenchmarkthinkingdefaultSource
GPQA Diamond–75.8%#131Epoch AI Benchmarking Hub
SimpleBench–36.1%#63SimpleBench
LMArena Hard Prompts–1426#123LMArena
GPQA Diamond (AA)not in index–69.3%#294Artificial Analysis
Humanity's Last Exam (AA)not in index–4.1%#455Artificial Analysis
GPQA Diamond (Vals)not in index–74.2%#82Vals AI
Kagi LLM Benchmark–61.3%#47Kagi LLM Benchmark
ARC-AGI-1–5.5%#174ARC Prize
ARC-AGI-2–0.0%#174ARC Prize

Coding

BenchmarkthinkingdefaultSource
WeirdML–37.2%#116WeirdML
LMArena Coding–1443#135LMArena
LiveCodeBench–52.9%#110Vals AI
Aider Polyglot–53.3%#21Aider polyglot leaderboard

Agents & tools

BenchmarkthinkingdefaultSource
APEX-Agents–2.1%#60Mercor
Terminal-Bench Hard–11.4%#207Artificial Analysis
τ²-Bench Telecom (AA)not in index–48.8%#194Artificial Analysis

Maths

BenchmarkthinkingdefaultSource
OTIS Mock AIME–55.6%#157Epoch AI Benchmarking Hub
MATH Level 5–88.8%#22Epoch AI Benchmarking Hub
AIME (Vals)–58.8%#57Vals AI
MGSM–91.3%#27Vals AI

Knowledge

BenchmarkthinkingdefaultSource
AA-Omniscience–-33.2#226Artificial Analysis
MMLU-Pro–80.0%#88Vals AI
LegalBench–82.6%#60Vals AI
CorpFin–59.7%#72Vals AI
TaxEval–75.9%#10Vals AI
MedQA–83.8%#61Vals AI

Instruction following

BenchmarkthinkingdefaultSource
IFBench–46.9%#189Artificial Analysis

Human preference

BenchmarkthinkingdefaultSource
LMArena Text–1412#119LMArena

Multimodal

BenchmarkthinkingdefaultSource
MMMU (validation)–78.0%#10MMMU

Long context

BenchmarkthinkingdefaultSource
Fiction.LiveBench 120k–58.3%#19Fiction.live
AA-LCR–58.0%#226Artificial Analysis

Composite

BenchmarkthinkingdefaultSource
Epoch Capabilities Indexnot in index–138.3#105Epoch AI Benchmarking Hub
AA Intelligence Index10.4#30512.1#270Artificial Analysis

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. Reasoning tokens are not modelled.

WorkloadTokens in / outCostWith cachingTime
Chat reply400 / 300$0.0013––
Summarise a 30-page report12,000 / 600$0.017––
Code edit6,000 / 1,500$0.011––
Agentic coding session60,000 / 4,000$0.085––
Structured extraction2,000 / 200$0.0030––

See also

Best under $2/MBest for codingBest for agentsAll xAI models

Data as of 9 Sept 2026. Compare these configurations.