Codestral 2508 vs R1 0528

Summary of our 12-test comparison (scores are our internal 1-5 ratings unless noted):

• Wins for R1 0528 (in our testing): strategic_analysis 4 vs 2, constrained_rewriting 4 vs 3, creative_problem_solving 4 vs 2, classification 4 vs 3, safety_calibration 4 vs 1, persona_consistency 5 vs 3, agentic_planning 5 vs 4, multilingual 5 vs 4. These wins indicate R1 is substantially better at nuanced tradeoff reasoning (strategic_analysis), refusing/allowing appropriately (safety_calibration), maintaining personas, goal decomposition and recovery (agentic_planning), and multilingual parity. R1's rankings reinforce this: safety_calibration rank 6 of 55 (tied with 3 others) and agentic_planning tied for 1st of 54.
• Wins for Codestral 2508 (in our testing): structured_output 5 vs 4. Codestral is superior at JSON/schema compliance and strict format adherence; it is tied for 1st on structured_output (tied with 24 other models out of 54). That makes Codestral the safer choice when exact format output matters (APIs, code generation where parsable output is required).
• Ties (in our testing): tool_calling (5/5), faithfulness (5/5), long_context (5/5). Both models score top marks on function selection/argument accuracy, sticking to source material, and retrieval across long contexts (30K+ tokens). Both are effectively equal for long-context retrieval and tool integration in our benchmarks.
• External benchmarks (Epoch AI): R1 0528 posts a math_level_5 score of 96.6% (rank 5 of 14 on that external test) and an AIME 2025 score of 66.4% (rank 16 of 23) — these third-party results complement our internal findings that R1 is strong on math/analysis tasks. Codestral has no external scores in the payload. Interpretation for real tasks: pick R1 when safety, planning, multilingual output, classification accuracy, or creative solutions matter; pick Codestral when you need guaranteed schema-compliant output and the lowest per-token cost for high-frequency coding flows like fill-in-the-middle, code correction, and test generation.

Costs per million input+output tokens (using per-mTok prices in the payload): Codestral 2508 = $0.30 (input) + $0.90 (output) = $1.20 per 1M input+output tokens. R1 0528 = $0.50 + $2.15 = $2.65 per 1M. At scale: 1M tokens/month costs $1.20 (Codestral) vs $2.65 (R1) — a $1.45 absolute difference. At 10M: $12.00 vs $26.50 (difference $14.50). At 100M: $120.00 vs $265.00 (difference $145.00). Teams with high-volume API usage (10M+ tokens/mo) or tight margins should prefer Codestral for cost efficiency; teams that need higher performance on safety, planning, multilingual, classification, or creative/problem-solving should budget for R1's ~2.2x per-token cost.

Choose Codestral 2508 if: you need the cheapest per-token model for high-frequency coding workflows, strict JSON/schema adherence (structured_output = 5), and low-latency code-focused tasks. Choose R1 0528 if: you require stronger safety calibration (4 vs 1), better agentic planning (5 vs 4), higher persona consistency and classification (5 vs 3/4), superior creative problem solving (4 vs 2), multilingual parity, or stronger external math results (96.6% on MATH Level 5, Epoch AI). If budget is the primary constraint at volumes ≥10M tokens/month, Codestral materially reduces spend; if capability and risk mitigation are primary, pay the premium for R1.