Devstral 2 2512 vs Gemini 3 Flash Preview

Summary of head-to-head results in our 12-test suite: Gemini 3 Flash Preview wins 7 tests, Devstral 2 2512 wins 1, and 4 tests tie. Where Gemini wins: strategic_analysis (5 vs 4) — Gemini ranks "tied for 1st" (rank 1 of 54); creative_problem_solving (5 vs 4) — Gemini ranks tied for 1st; tool_calling (5 vs 4) — Gemini is tied for 1st (display: "tied for 1st with 16 other models"); faithfulness (5 vs 4) — Gemini ties for 1st (rank 1 of 55); classification (4 vs 3) — Gemini ties for 1st (display: "tied for 1st with 29 other models"); persona_consistency (5 vs 4) — Gemini ties for 1st; agentic_planning (5 vs 4) — Gemini ties for 1st. Devstral's win: constrained_rewriting (5 vs 4) — Devstral is "tied for 1st with 4 other models out of 53 tested," which matters for strict-length compression or tight character-limit tasks. Ties: structured_output (5 each), long_context (5 each), multilingual (5 each), and safety_calibration (both 1). Practical meaning: Gemini's higher scores and top-tier ranks on tool_calling, agentic_planning, and faithfulness translate to more accurate function selection, stronger multi-step goal decomposition, and fewer deviations from source material in our tests. Devstral matches Gemini on long-context and structured output and outperforms it on constrained rewriting, so it's preferable for tasks requiring tight output compression. External benchmarks (supplementary): Gemini 3 Flash Preview scores 75.4% on SWE-bench Verified (Epoch AI) and 92.8% on AIME 2025 (Epoch AI), which supports its coding and high-level math performance; Devstral has no external scores in the payload. Note both models score 1 on safety_calibration and rank 32 of 55 in our tests — neither reliably refuses disallowed or harmful prompts in our suite.

Per-million-token pricing (input/output per payload): Devstral 2 2512 charges $0.40 input / $2.00 output; Gemini 3 Flash Preview charges $0.50 input / $3.00 output. Assuming a 50/50 split of input/output tokens, monthly costs are: 1M tokens — Devstral $1.20 vs Gemini $1.75 (Gemini +$0.55, +45.8%); 10M tokens — Devstral $12.00 vs Gemini $17.50 (Gemini +$5.50); 100M tokens — Devstral $120.00 vs Gemini $175.00 (Gemini +$55.00). If your workload is output-heavy (e.g., mostly generation), the gap is larger: at 1M output-only tokens Devstral is $2.00 vs Gemini $3.00 (+$1/M). Teams generating tens of millions of tokens should care: the difference scales linearly (Gemini costs roughly 1.46x more under a balanced split).

Choose Devstral 2 2512 if: you need a lower-cost option (≈33% cheaper per-M under balanced usage), you require strong constrained_rewriting or top-tier structured output at a 256K context window, and you prioritize cost-efficiency at scale. Choose Gemini 3 Flash Preview if: you need best-in-suite tool calling, agentic planning, classification, faithfulness, and creative/problem-solving performance (it wins 7 of 12 benchmarks and holds high ranks), require multimodal inputs or very large context (1,048,576), and can justify the ~1.46x cost under balanced token usage.