BenchLeader
Menu
LeaderboardsModelsProvidersBenchmarksCompareReleasesBlogMethodology
Mistral AIOpen weightsReasoning modelAuto-detected

Magistral Small

Best configuration ranks #504 of 610 on the BenchLeader Index at 40.5 ±5.7. Last measured 1 Sept 2026. Released 18 Sept 2025.

Blended price
$0.750/M
$0.500 in · $1.50 out
Output speed
–
First answer
–
Context
128k
How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.
  1. Overall index41
  2. Reasoning31
  3. Coding46
  4. Agents & tools38
  5. Maths46
  6. Knowledge30
  7. Instruction following46
  8. Multimodal40
  9. Long context35
  10. Composite40

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

BenchmarkdefaultSource
GPQA Diamond56.1%#186Epoch AI Benchmarking Hub
GPQA Diamond (AA)not in index66.3%#319Artificial Analysis
Humanity's Last Exam (AA)not in index7.5%#291Artificial Analysis
GPQA Diamond (Vals)not in index58.3%#108Vals AI
Kagi LLM Benchmark6.3%#135Kagi LLM Benchmark
ARC-AGI-15.0%#178ARC Prize
ARC-AGI-20.0%#174ARC Prize

Coding

BenchmarkdefaultSource
SciCode35.2%#135SciCode
LiveCodeBench72.1%#81Vals AI

Agents & tools

BenchmarkdefaultSource
Terminal-Bench Hard4.5%#270Artificial Analysis
τ²-Bench Telecom (AA)not in index27.8%#264Artificial Analysis

Maths

BenchmarkdefaultSource
OTIS Mock AIME30.0%#190Epoch AI Benchmarking Hub
AIME (Vals)80.7%#47Vals AI
MGSM86.3%#60Vals AI

Knowledge

BenchmarkdefaultSource
AA-Omniscience-44.4#282Artificial Analysis
MMLU-Pro62.1%#126Vals AI
LegalBench40.0%#135Vals AI
CorpFin44.0%#113Vals AI
TaxEval60.3%#123Vals AI
MedQA82.4%#64Vals AI

Instruction following

BenchmarkdefaultSource
IFBench44.4%#208Artificial Analysis

Multimodal

BenchmarkdefaultSource
MMMU-Pro55.5%#192Artificial Analysis

Long context

BenchmarkdefaultSource
AA-LCR19.3%#366Artificial Analysis

Composite

BenchmarkdefaultSource
AA Intelligence Index8.6#357Artificial 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$0.0007––
Summarise a 30-page report12,000 / 600$0.0069––
Code edit6,000 / 1,500$0.0053––
Agentic coding session60,000 / 4,000$0.036––
Structured extraction2,000 / 200$0.0013––

See also

Best under $2/MBest open weightsBest for codingAll Mistral AI models

Data as of 9 Sept 2026. Compare with another model.