Codestral 2508 vs DeepSeek V3.1 Terminus

Summary of our 12-test comparison (scores on a 1–5 scale):

• Codestral 2508 wins: tool_calling 5 vs 3, faithfulness 5 vs 3. In our testing Codestral is tied for 1st in faithfulness ("tied for 1st with 32 other models out of 55 tested") and tied for 1st on tool_calling ("tied for 1st with 16 other models out of 54 tested"). That indicates stronger function selection, argument accuracy, sequencing and strict adherence to source — important for code generation, refactors, and automated tool use.
• DeepSeek V3.1 Terminus wins: strategic_analysis 5 vs 2, creative_problem_solving 4 vs 2, persona_consistency 4 vs 3, multilingual 5 vs 4. Notably, DeepSeek's strategic_analysis is "tied for 1st with 25 other models out of 54 tested," and creative_problem_solving ranks 9 of 54 (tied with 20). This shows DeepSeek is meaningfully better at nuanced tradeoff reasoning, non-obvious solution generation, maintaining persona, and non-English outputs.
• Ties (no clear winner): structured_output 5/5 (both tied for 1st with 24 others), constrained_rewriting 3/3, classification 3/3, long_context 5/5 (both tied for 1st with 36 others), safety_calibration 1/1, agentic_planning 4/4. Ties on structured_output and long_context mean both handle JSON/schema compliance and 30K+ contexts at the top end of tested models. Low safety_calibration (1) for both flags identical caution: neither model excels at sensitive refusal/allow decisions in our tests. What this means for real tasks: choose Codestral for deterministic codegen, tool-driven pipelines, and tasks demanding high faithfulness to source code or specs. Choose DeepSeek for tasks requiring multi-step reasoning, creative solutions, persona fidelity, or multilingual output. Both tie on structured output and long context, so both are viable where large contexts or strict formats matter.

Costs are per million tokens. Codestral 2508: $0.30 (input) / $0.90 (output) per M. DeepSeek V3.1 Terminus: $0.21 (input) / $0.79 (output) per M. If usage is balanced 50/50 input/output, per-M costs are $0.60 (Codestral) vs $0.50 (DeepSeek). At 1M balanced tokens/month you save $0.10 with DeepSeek; at 10M you save $1.00; at 100M you save $10.00. If your workload is output-heavy (e.g., large generations), Codestral costs $0.90 vs DeepSeek $0.79 per M — $0.11/M difference ($11 at 100M). Teams running high-volume inference (10M–100M tokens/month) should care about the small but accumulating gap; low-volume users (<1M/month) will see negligible monthly difference.

Choose Codestral 2508 if you prioritize coding accuracy, tool calling, and source-faithful code (tool_calling 5, faithfulness 5) and you accept ~14% higher per-token cost. Choose DeepSeek V3.1 Terminus if you need stronger strategic analysis, creative problem solving, persona consistency, or multilingual output (strategic_analysis 5, creative_problem_solving 4, multilingual 5) and want lower per-token costs.