Codestral 2508 vs Grok 3 Mini

Summary of our 12-test comparison (scores are from our testing):

• Ties (both models): tool_calling 5/5 (tied for 1st with 16 others), faithfulness 5/5 (tied for 1st with 32 others), long_context 5/5 (tied for 1st with 36 others), multilingual 4/4 (tie). Practically: both handle function selection, long contexts (30K+), and faithfulness equally well in our benchmarks.
• Codestral 2508 wins: structured_output 5 vs 4 (Codestral tied for 1st with 24 others; Grok rank 26/54). This matters for JSON/schema-compliant code and tools that require exact format adherence. Agentic_planning 4 vs 3 (Codestral rank 16/54 vs Grok rank 42/54), meaning Codestral did better at goal decomposition and recovery in our tests.
• Grok 3 Mini wins: persona_consistency 5 vs 3 (Grok tied for 1st with 36 others; Codestral rank 45/53) — Grok resists injection and keeps character better. Classification 4 vs 3 (Grok tied for 1st with 29 others; Codestral rank 31/53) — Grok is stronger at routing and categorization in our suite. Constrained_rewriting 4 vs 3 (Grok rank 6/53; Codestral rank 31/53) — Grok compresses into hard limits more reliably. Creative_problem_solving 3 vs 2 (Grok rank 30/54; Codestral rank 47/54) and strategic_analysis 3 vs 2 (Grok rank 36/54; Codestral rank 44/54) — Grok produced more feasible non-obvious ideas and nuanced tradeoffs in our tests. Safety_calibration 2 vs 1 (Grok rank 12/55; Codestral rank 32/55) — Grok refused harmful prompts more appropriately in our benchmarks.
• Interpretation for real tasks: if your priority is exact structured outputs (APIs, code snippets, lintable JSON) and stronger multi-step planning for automation, Codestral 2508 shows the edge. If you need lower cost, better persona consistency, safer refusals, stronger classification/routing, or better constrained rewriting, Grok 3 Mini wins in our testing and at lower output cost.

Both models share the same input price ($0.30/mTok) but Codestral 2508 charges $0.90/mTok for outputs vs Grok 3 Mini at $0.50/mTok (price ratio 1.8). Practical cost examples (balanced I/O split 50/50): Codestral = $0.60 per 1M tokens → $6.00 per 10M → $60.00 per 100M; Grok 3 Mini = $0.40 per 1M → $4.00 per 10M → $40.00 per 100M. Write-heavy workloads (90% output): Codestral = $0.84/1M → $8.40/10M → $84/100M; Grok 3 Mini = $0.48/1M → $4.80/10M → $48/100M. Read-heavy (10% output): Codestral = $0.36/1M → $3.60/10M → $36/100M; Grok 3 Mini = $0.32/1M → $3.20/10M → $32/100M. Who should care: teams generating large volumes of output tokens (code generation, long-form content) will see meaningful savings with Grok 3 Mini; developer desks or small experiments may prioritize Codestral 2508's structured-output and agentic strengths despite the higher per-output cost.

Choose Codestral 2508 if: you prioritize the highest structured-output fidelity and stronger agentic planning in code-heavy workflows (structured_output 5 vs 4; agentic_planning 4 vs 3), and you accept higher output costs ($0.90/mTok). Choose Grok 3 Mini if: you need a lower-cost model with stronger safety, persona consistency, classification, and constrained rewriting (Grok wins 6 of 12 benchmarks in our tests), or you operate at scale where output-cost savings compound.