Codestral 2508 vs R1

Head-to-head wins (per our 12-test suite): Codestral 2508 wins 4 tests, R1 wins 5, and 3 tests tie. Detailed walk-through:

• Structured output: Codestral 2508 wins (score 5 vs R1 4). This matters for strict JSON/schema outputs; Codestral is tied for 1st with 24 others on structured_output in our rankings ("tied for 1st with 24 other models out of 54 tested").
• Tool calling: Codestral 2508 wins (5 vs 4). Codestral is tied for 1st in tool_calling in our rankings ("tied for 1st with 16 other models out of 54 tested"), so it will select functions, arguments and sequencing more reliably in our tests.
• Classification: Codestral 2508 wins (3 vs 2). Codestral ranks 31 of 53 on classification while R1 ranks 51 of 53; expect fewer routing/mapping errors with Codestral in our classification probe.
• Long context: Codestral 2508 wins (5 vs 4). Codestral is tied for 1st on long_context ("tied for 1st with 36 other models out of 55 tested"), so retrieval and accuracy across 30K+ tokens favors Codestral.
• Strategic analysis: R1 wins decisively (5 vs 2). R1 is tied for 1st on strategic_analysis in our rankings ("tied for 1st with 25 other models out of 54 tested"), meaning nuanced tradeoff reasoning with real numbers was substantially better in our tests.
• Constrained rewriting: R1 wins (4 vs 3). R1 ranks 6 of 53 in constrained_rewriting, so it compresses content into hard character limits more effectively.
• Creative problem solving: R1 wins (5 vs 2). R1 is tied for 1st on creative_problem_solving, producing more non-obvious, specific feasible ideas in our tasks.
• Persona consistency: R1 wins (5 vs 3). R1 is tied for 1st on persona_consistency ("tied for 1st with 36 other models"), resisting injection and maintaining character better in our tests.
• Multilingual: R1 wins (5 vs 4). R1 is tied for 1st in multilingual quality across our languages, while Codestral ranks mid-pack.
• Faithfulness: tie (both 5). Both models tied for 1st on faithfulness ("tied for 1st with 32 other models out of 55 tested"), so both stick to source material in our evaluation.
• Safety calibration: tie (both 1). Neither model scored well in safety_calibration in our tests (both low and both rank 32 of 55), so expect cautious evaluation and safety testing regardless of choice.
• Agentic planning: tie (both 4). Both models rank similarly on agentic_planning (rank 16 of 54) meaning similar decomposition and failure-recovery capability in our suite. External benchmarks: R1 includes third-party math results in the payload: math_level_5 = 93.1% and aime_2025 = 53.3% (Epoch AI). Codestral 2508 has no external math scores in the payload. Use these Epoch AI numbers as supplementary signals for R1's strength on high-difficulty math tasks.

Per-token rates (input/output per mTok): Codestral 2508 = $0.30 / $0.90; R1 = $0.70 / $2.50. The payload gives a priceRatio of 0.36 (Codestral costs ~36% of R1 for the same token mix). Cost scenarios for 1M / 10M / 100M tokens (1M tokens = 1,000 mTok):

• Codestral 2508: • Input-only: 1M = $300; 10M = $3,000; 100M = $30,000 • Output-only: 1M = $900; 10M = $9,000; 100M = $90,000 • 50/50 input/output (illustrative): 1M = $600; 10M = $6,000; 100M = $60,000
• R1: • Input-only: 1M = $700; 10M = $7,000; 100M = $70,000 • Output-only: 1M = $2,500; 10M = $25,000; 100M = $250,000 • 50/50 input/output (illustrative): 1M = $1,600; 10M = $16,000; 100M = $160,000 Who should care: teams at scale (10M+ tokens/month) will see the gap magnify into tens or hundreds of thousands of dollars. Cost-sensitive deployments (large-scale assistants, CI coding jobs, automated test generation) will favor Codestral 2508; R1 is costlier but may justify the spend where superior strategic/creative reasoning is critical.

Choose Codestral 2508 if: you need low-latency, cost-efficient production for coding workflows, function/tool calling, strict structured outputs or very long-context retrieval — it wins tool_calling, structured_output and long_context in our tests and costs far less ($0.30/$0.90 per mTok). Choose R1 if: you prioritize strategic reasoning, creative problem solving, persona consistency or multilingual excellence — R1 wins those benchmarks (strategic_analysis, creative_problem_solving, persona_consistency, multilingual) and posts strong external math scores (math_level_5 93.1% and AIME 2025 53.3% per Epoch AI), but expect significantly higher costs.