Codestral (25.08) vs GPT-5.4

Best cell per row highlighted. Null means undisclosed — never counted as zero.

Mistral AI
Codestral (25.08)GA
7.6
AI Panel
OpenAI
GPT-5.4GA
8.7
AI Panel

Identity & lifecycle

ProviderMistral AIOpenAI
Family / tierCodestral · CoderGPT-5 · Medium
StatusGAGA
Released2025-07-312026-03-05
Knowledge cutoff2025-052025-08

Architecture & context

Context window128K1.1M
Max output tokens16K128K
Input modalitiestexttext, image
Reasoning modenoneoptional
Open weightsNoNo

Pricing (per Mtok)

Input$0.3$2.5
Output$0.9$15
Cached input$0.03$0.25
Batch input$0.15$1.25
Free tierYesNo

Speed

Output speed (tok/s)—86.3
Time to first token (s)—1.2
Latency tierfastmedium

Trust & deployment

Trains on inputsNoNo
LicenseProprietaryProprietary
CloudsBedrock, Azure AI Foundry, Vertex AIAzure OpenAI, Azure AI Foundry
ComplianceSOC2, ISO27001, GDPRSOC2, GDPR, CCPA, HIPAA

AI Panel scoring

Unified score7.68.7
Decision Maker7.59
Domain Strategist7.58.7
Finance Lead89
Domain Practitioner89
Power User7.58.5
Skeptic78
Value score8.58.5

Benchmarks

GPQA Diamond—92.8
HumanEval86.6—
MMMU—81.2
Terminal-Bench—75.1
MRCR Long Context—36.6
Humanity's Last Exam—39.8
LMArena Elo—1476
LMArena Coding Elo—1520

Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.