DeepSeek V3.2 vs Devstral Small 1.1

Across our 12-test suite, DeepSeek V3.2 outscores Devstral Small 1.1 on 9 benchmarks, loses 2, and ties 1.

Where DeepSeek V3.2 wins clearly:

• Agentic planning (5 vs 2): DeepSeek V3.2 ties for 1st among 54 models (with 14 others); Devstral Small 1.1 ranks 53rd of 54. This is the widest gap in the comparison and matters most for multi-step autonomous workflows — goal decomposition, failure recovery, and sequential decision-making.
• Strategic analysis (5 vs 2): DeepSeek V3.2 again ties for 1st (26 models share this score) vs Devstral Small 1.1 at rank 44 of 54. For nuanced tradeoff reasoning with real numbers, DeepSeek V3.2 is in a different league.
• Creative problem solving (4 vs 2): DeepSeek V3.2 ranks 9th of 54; Devstral Small 1.1 ranks 47th of 54. Non-obvious ideation and lateral thinking heavily favor DeepSeek V3.2.
• Persona consistency (5 vs 2): DeepSeek V3.2 ties for 1st (37 models); Devstral Small 1.1 ranks 51st of 53 — near the bottom. For chat products, roleplay systems, or any application requiring stable character under adversarial prompts, this is a significant concern for Devstral Small 1.1.
• Faithfulness (5 vs 4): DeepSeek V3.2 ties for 1st (33 models); Devstral Small 1.1 ranks 34th of 55. Both are decent, but DeepSeek V3.2 is more reliable at sticking to source material without hallucinating.
• Multilingual (5 vs 4): DeepSeek V3.2 ties for 1st (35 models); Devstral Small 1.1 ranks 36th of 55. Both produce non-English output, but DeepSeek V3.2 matches the field leaders.
• Structured output (5 vs 4): DeepSeek V3.2 ties for 1st (25 models); Devstral Small 1.1 ranks 26th of 54. JSON schema compliance is strong on both, but DeepSeek V3.2 is more consistent.
• Long context (5 vs 4): DeepSeek V3.2 ties for 1st (37 models); Devstral Small 1.1 ranks 38th of 55. DeepSeek V3.2 also has a larger context window (163,840 vs 131,072 tokens).
• Constrained rewriting (4 vs 3): DeepSeek V3.2 ranks 6th of 53; Devstral Small 1.1 ranks 31st of 53.

Where Devstral Small 1.1 wins:

• Tool calling (4 vs 3): Devstral Small 1.1 ranks 18th of 54; DeepSeek V3.2 ranks 47th of 54. This is a meaningful gap for function-dispatch workloads — Devstral Small 1.1 is more accurate on function selection, argument construction, and sequencing. Notably, the median model in our suite scores 4 on tool calling, so DeepSeek V3.2's 3/5 is below average here.
• Classification (4 vs 3): Devstral Small 1.1 ties for 1st (30 models); DeepSeek V3.2 ranks 31st of 53. For routing, tagging, and categorization tasks, Devstral Small 1.1 is the better pick.

Tie:

• Safety calibration (2 vs 2): Both models score 2/5, ranking 12th of 55 (tied with 19 others). Safety calibration scores are low across the board — the field median is 2 — so neither model stands out here.

DeepSeek V3.2 costs $0.26/M input and $0.38/M output. Devstral Small 1.1 costs $0.10/M input and $0.30/M output — making it 2.6× cheaper on input and about 27% cheaper on output. In practice: at 1M output tokens/month, that's $0.38 vs $0.30 — an $0.08 difference that's effectively noise. At 10M output tokens, the gap is $0.80 — still minor. At 100M output tokens/month, you're looking at $38,000 vs $30,000 — an $8,000 annual difference that may matter at scale. The input cost gap is more significant for retrieval-heavy or long-context workloads: DeepSeek V3.2's 163,840-token context window is larger, but feeding large contexts through it at $0.26/M vs $0.10/M adds up faster. Teams running high-volume classification or tool-dispatch pipelines — the two areas where Devstral Small 1.1 wins — could justify the cheaper model. Everyone else is paying a small premium for substantially better reasoning, planning, and multilingual output.

Choose DeepSeek V3.2 if you're building agentic systems, pipelines requiring multi-step planning, or anything that involves strategic reasoning, long documents, or non-English languages. It scores 5/5 on agentic planning, strategic analysis, long context, and multilingual in our testing — each at or near the top of our 52+ model pool. It's also the better choice for applications that need stable persona behavior or high faithfulness to source material. The $0.08/M output cost premium over Devstral Small 1.1 is negligible for most use cases.

Choose Devstral Small 1.1 if your workload is primarily tool calling or classification — the two benchmarks where it outperforms DeepSeek V3.2 (scoring 4 vs 3 on both). It's purpose-built for software engineering agent contexts and ranks 18th of 54 on tool calling vs DeepSeek V3.2's 47th. At $0.10/M input, it's also the cheaper option for high-volume classification or routing pipelines where you don't need broad reasoning capabilities. If you're running >100M output tokens/month and the workload is narrowly tool-dispatch or categorization, the cost difference ($8,000/month at that volume) becomes a real factor.