Claude Fable 5 vs Qwen2.5-Coder-32B-Instruct vs Codestral (25.08)

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

Anthropic
Claude Fable 5GA
9.5
AI Panel
Alibaba Cloud
Qwen2.5-Coder-32B-InstructGA
8.4
AI Panel
Mistral AI
Codestral (25.08)GA
7.6
AI Panel

Identity & lifecycle

ProviderAnthropicAlibaba CloudMistral AI
Family / tierClaude 5 · FableQwen2.5-Coder · CoderCodestral · Coder
StatusGAGAGA
Released2026-06-092024-11-122025-07-31
Knowledge cutoff2026-012024-082025-05

Architecture & context

Context window1M131K128K
Max output tokens128K8K16K
Input modalitiestext, image, filetexttext
Reasoning modealwaysnonenone
Open weightsNoYesNo

Pricing (per Mtok)

Input$10$0.08$0.3
Output$50$0.24$0.9
Cached input$1—$0.03
Batch input$5—$0.15
Free tierNoYesYes

Speed

Output speed (tok/s)———
Time to first token (s)———
Latency tierslowfastfast

Trust & deployment

Trains on inputsNoNoNo
LicenseProprietaryApache-2.0Proprietary
CloudsBedrock, Vertex AI, Azure AI FoundryGCPBedrock, Azure AI Foundry, Vertex AI
ComplianceSOC2, HIPAA, GDPR, ISO27001—SOC2, ISO27001, GDPR

AI Panel scoring

Unified score9.58.47.6
Decision Maker9.58.57.5
Domain Strategist9.587.5
Finance Lead798
Domain Practitioner9.59.58
Power User98.57.5
Skeptic7.57.57
Value score6.59.58.5

Benchmarks

SWE-bench Verified95——
HumanEval—92.786.6
LiveCodeBench—31.4—
Terminal-Bench88——
LMArena Elo15061270—
LMArena Coding Elo15521342—

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