Claude Haiku 4.5 vs Codestral 2508 for Classification

What Classification demands: fast, repeatable mapping of inputs to categories or routes with reliable, schema-compliant outputs and safe refusal behavior. Key capabilities: 1) classification accuracy (our classification test is the primary measure), 2) structured output (JSON/schema adherence), 3) faithfulness and context sensitivity (to avoid misrouting), 4) safety calibration (correctly refuse harmful labels), and 5) cost/latency tradeoffs for high-volume inference. Primary signal: our Classification score and task rank — Claude Haiku 4.5 scores 4 (taskRank 1/52) vs Codestral 2508 at 3 (taskRank 31/52). Supporting evidence from our proxies: Haiku leads on safety_calibration (2 vs 1), persona_consistency (5 vs 3), and agentic_planning (5 vs 4), all helpful when classification depends on context or multi-step interpretation. Codestral leads on structured_output (5 vs 4), making it advantageous when strict JSON/schema adherence is essential. Tool_calling, faithfulness, and long_context are ties (both models score 5) and will not differentiate basic routing workflows.

1. Customer-support routing: Haiku 4.5 (score 4) — preferable when intent is subtle, context matters across a long conversation, and safety/policy-aware refusals are required (safety_calibration 2 vs 1). 2) High-throughput labeler with strict JSON schema: Codestral 2508 — better structured output (5 vs 4) and much lower output cost (0.9 vs 5 per mTok), reducing inference bill while enforcing schema. 3) Moderation triage: Haiku 4.5 reduces risky misclassifications because persona_consistency and safety are stronger; both models tie on faithfulness (5) so neither is prone to inventing sources. 4) Embedded device or edge batch classification where price matters: Codestral is ~5.56x cheaper per output mTok (priceRatio 5.555...), making it the cost-efficient option if a 1-point drop in our classification score is acceptable. 5) Multi-turn classification that needs tool orchestration: both models score 5 on tool_calling, so either can select and sequence tools reliably; prefer Haiku if context-sensitive decisions matter, prefer Codestral if schema cost constraints dominate.

For Classification, choose Claude Haiku 4.5 if you need higher accuracy in subtle, context-heavy routing, better safety calibration, and stronger persona/context handling (score 4 vs 3, rank 1 vs 31). Choose Codestral 2508 if you require top-tier structured output (5 vs 4) and much lower inference cost (output cost per mTok 0.9 vs 5), and you can accept a 1-point drop on our Classification test.