Devstral 2 2512 vs Devstral Medium

We ran both models across our 12-test suite and compare scores (1-5) and rankings below — all claims are from our testing. Summary from win/tie data: Devstral 2 2512 wins 8 categories, Devstral Medium wins 1, and 3 are ties. Detailed walk‑through:

• Structured output: Devstral 2 2512 scores 5 vs Devstral Medium 4. In our testing Devstral 2 2512 is tied for 1st ("tied for 1st with 24 other models out of 54 tested"), while Devstral Medium sits at rank 26 of 54. This matters when you must produce strict JSON/schema-compliant responses.

• Constrained rewriting: Devstral 2 2512 scores 5 vs Devstral Medium 3. Devstral 2 2512 is tied for 1st ("tied for 1st with 4 other models out of 53 tested"). For tasks requiring tight character limits or compression, Devstral 2 2512 is objectively better in our benchmarks.

• Long context: Devstral 2 2512 scores 5 vs Devstral Medium 4; Devstral 2 2512 is tied for 1st ("tied for 1st with 36 other models out of 55 tested"). With a 262,144 token window vs 131,072, Devstral 2 2512 performs better at retrieval/understanding across 30K+ token inputs in our tests.

• Tool calling: Devstral 2 2512 scores 4 vs Devstral Medium 3. Rankings show Devstral 2 2512 at rank 18 of 54 and Devstral Medium at rank 47 of 54. This indicates more accurate function selection and argument sequencing for Devstral 2 2512 in our tool-calling tasks.

• Creative problem solving: Devstral 2 2512 scores 4 vs Devstral Medium 2 (rank 9 vs rank 47 in their respective full distributions). Devstral 2 2512 produced more feasible, specific ideas in our creative tasks.

• Strategic analysis: Devstral 2 2512 scores 4 vs Devstral Medium 2; Devstral 2 2512 ranks 27 of 54 vs Devstral Medium 44 of 54. For nuanced numeric tradeoffs and multi-step reasoning, Devstral 2 2512 is stronger in our benchmarks.

• Persona consistency and multilingual: Devstral 2 2512 scored 4 and 5 vs Devstral Medium's 3 and 4 respectively; Devstral 2 2512 ties for top ranks in multilingual tasks ("tied for 1st with 34 other models out of 55 tested"). Expect better non-English parity and resistance to prompt injection on persona tasks with Devstral 2 2512 in our tests.

• Classification: Devstral Medium wins here (score 4 vs Devstral 2 2512's 3). Devstral Medium is tied for 1st on classification ("tied for 1st with 29 other models out of 53 tested"), so it is preferable when routing/categorization accuracy is the primary requirement.

• Faithfulness, safety_calibration, agentic_planning: these are ties (both models score equally at 4/1/4 respectively in our tests). For hallucination resistance, refusal behavior, and high-level goal decomposition, neither model had a clear advantage on our benchmarks.

Practical takeaway: Devstral 2 2512 is measurably stronger for strict schema outputs, long-context retrieval, constrained rewriting, tool orchestration and creative problem solving in our testing; Devstral Medium is narrowly better for classification tasks.

Both models have identical pricing: input_cost_per_mtok = $0.40 and output_cost_per_mtok = $2.00. That equals $400 per 1M input tokens and $2,000 per 1M output tokens. If you assume a 50/50 split of input/output tokens, cost totals are: 1M combined tokens ≈ $1,200/month, 10M ≈ $12,000/month, 100M ≈ $120,000/month. Because output tokens dominate cost ($2.00 vs $0.40), teams optimizing spend should focus on reducing generated output length or caching; there is no price gap between Devstral 2 2512 and Devstral Medium to drive the decision.

Choose Devstral 2 2512 if you need: strict JSON/schema compliance, long-context handling (262,144 token window), constrained rewriting, stronger tool-calling and agentic coding workflows, or better multilingual output — it won 8 of 12 benchmarks in our tests. Choose Devstral Medium if your priority is classification/routing (it wins classification, score 4 vs 3) or you need a smaller context window (131,072) but otherwise want the same pricing and parameter support. Because pricing is identical, pick by capability: pick Devstral 2 2512 for developer tooling, long-context apps, and structured outputs; pick Devstral Medium when classification accuracy is the dominant metric.