DeepSeek V3.2 vs Devstral Medium

Across our 12-test benchmark suite, DeepSeek V3.2 wins 10 categories, Devstral Medium wins 1 (classification), and they tie on tool calling. Here's the breakdown:

Where DeepSeek V3.2 leads:

• Strategic analysis (5 vs 2): DeepSeek V3.2 ties for 1st among 54 models tested; Devstral Medium sits at rank 44 of 54. That's a 3-point gap on a 5-point scale — a decisive difference for financial modeling, tradeoff reasoning, and analytical writing.
• Creative problem solving (4 vs 2): DeepSeek V3.2 ranks 9th of 54; Devstral Medium ranks 47th. For ideation, non-obvious solutions, and lateral thinking tasks, this gap is meaningful.
• Persona consistency (5 vs 3): DeepSeek V3.2 ties for 1st among 53 models; Devstral Medium ranks 45th. Critical for chatbot and assistant applications that need stable character and resistance to prompt injection.
• Faithfulness (5 vs 4): DeepSeek V3.2 ties for 1st among 55 models; Devstral Medium ranks 34th. For RAG pipelines and summarization, sticking to source material without hallucinating is a safety-critical property.
• Agentic planning (5 vs 4): DeepSeek V3.2 ties for 1st among 54 models (alongside 14 others); Devstral Medium ranks 16th. The gap is one point, but at the top of the distribution — goal decomposition and failure recovery both matter in multi-step autonomous workflows.
• Structured output (5 vs 4): DeepSeek V3.2 ties for 1st among 54 models; Devstral Medium ranks 26th. JSON schema compliance is table stakes for API integrations — DeepSeek V3.2 is more reliable here.
• Long context (5 vs 4): DeepSeek V3.2 ties for 1st among 55 models; Devstral Medium ranks 38th. Combined with its larger 163,840-token context window, DeepSeek V3.2 is better equipped for document-heavy tasks.
• Multilingual (5 vs 4): DeepSeek V3.2 ties for 1st among 55 models; Devstral Medium ranks 36th. For non-English deployments, the difference matters.
• Constrained rewriting (4 vs 3): DeepSeek V3.2 ranks 6th of 53; Devstral Medium ranks 31st. Compression within hard character limits is relevant for content pipelines, notifications, and ad copy.
• Safety calibration (2 vs 1): Both models score below the field median (p50 = 2), but DeepSeek V3.2 at rank 12 of 55 is meaningfully better than Devstral Medium at rank 32 of 55. Neither should be deployed in high-stakes safety contexts without additional guardrails.

Where Devstral Medium leads:

• Classification (4 vs 3): Devstral Medium ties for 1st among 53 models; DeepSeek V3.2 ranks 31st. For routing, tagging, and categorization pipelines, Devstral Medium has a genuine edge. This is its only benchmark win.

Tie:

• Tool calling (3 vs 3): Both models rank 47th of 54 — both are below the field median (p50 = 4). Neither model is a strong pick if tool calling accuracy is your primary requirement; you'd want to look elsewhere in the model landscape for that workload.

The pricing gap here is substantial and lopsided. DeepSeek V3.2 runs at $0.26/MTok input and $0.38/MTok output. Devstral Medium costs $0.40/MTok input and $2.00/MTok output — making output tokens more than 5x more expensive. In practice, output costs dominate most production workloads. At 1M output tokens/month, DeepSeek V3.2 costs $0.38 versus Devstral Medium's $2.00 — a $1.62 difference that's almost negligible. Scale to 10M tokens and that gap becomes $16.20/month. At 100M output tokens/month — a realistic volume for any production agentic pipeline — you're paying $380 with DeepSeek V3.2 versus $2,000 with Devstral Medium, a $1,620/month difference. For high-throughput applications like automated code review, document processing, or agentic task loops, that cost differential compounds fast. Devstral Medium also has a narrower context window (131,072 tokens vs DeepSeek V3.2's 163,840), so you get less capacity for more money. The only team that should seriously consider Devstral Medium's pricing is one where classification is the dominant task and the accuracy difference on that single benchmark justifies the 5x output cost.

Choose DeepSeek V3.2 if you need a general-purpose model that excels at analysis, agentic workflows, long-document tasks, multilingual output, or structured data generation. It wins 10 of 12 benchmarks in our testing and costs 81% less on output tokens ($0.38 vs $2.00/MTok) — making it the dominant choice for nearly every production use case, especially at scale. Its 163,840-token context window also gives it an edge in document-heavy or multi-turn applications.

Choose Devstral Medium if classification is your primary, isolated workload — routing emails, tagging tickets, categorizing content — and you specifically need the accuracy edge it shows in our testing (4 vs 3). Be aware that you'll pay 5x more per output token for that one-category advantage, and Devstral Medium scores below DeepSeek V3.2 on every other dimension we tested.