Devstral Medium vs Gemini 3.1 Flash Lite Preview

Across our 12-test suite, Gemini 3.1 Flash Lite Preview wins 9 benchmarks, Devstral Medium wins 1, and 2 are tied.

Where Gemini 3.1 Flash Lite Preview wins clearly:

• Safety calibration: 5 vs 1. This is the largest margin in the comparison. Gemini 3.1 Flash Lite Preview ties for 1st among 55 models tested; Devstral Medium ranks 32nd of 55. For applications where the AI must correctly refuse harmful requests while permitting legitimate ones — content moderation tools, consumer-facing products — this gap is a significant operational risk differentiator.
• Strategic analysis: 5 vs 2. Gemini 3.1 Flash Lite Preview ties for 1st of 54 models; Devstral Medium ranks 44th of 54. This test covers nuanced tradeoff reasoning with real numbers — business analysis, decision support, financial summaries.
• Persona consistency: 5 vs 3. Gemini 3.1 Flash Lite Preview ties for 1st of 53 models; Devstral Medium ranks 45th of 53. Relevant for chatbot personas and roleplay applications where character coherence matters.
• Multilingual: 5 vs 4. Gemini 3.1 Flash Lite Preview ties for 1st of 55 models; Devstral Medium ranks 36th. For non-English deployments, Gemini 3.1 Flash Lite Preview holds a meaningful edge.
• Structured output: 5 vs 4. Gemini 3.1 Flash Lite Preview ties for 1st of 54 models; Devstral Medium ranks 26th. JSON schema compliance and format adherence — critical for API pipelines.
• Faithfulness: 5 vs 4. Gemini 3.1 Flash Lite Preview ties for 1st of 55 models; Devstral Medium ranks 34th. Staying close to source material without hallucinating is essential for summarization and RAG use cases.
• Tool calling: 4 vs 3. Gemini 3.1 Flash Lite Preview ranks 18th of 54; Devstral Medium ranks 47th of 54. Notably, Devstral Medium's score places it near the bottom of the field on function selection and argument accuracy — a significant weakness for agentic workflows.
• Creative problem solving: 4 vs 2. Gemini 3.1 Flash Lite Preview ranks 9th of 54; Devstral Medium ranks 47th of 54. Devstral Medium's score here is near-floor performance on generating non-obvious, feasible ideas.
• Constrained rewriting: 4 vs 3. Gemini 3.1 Flash Lite Preview ranks 6th of 53; Devstral Medium ranks 31st.

Where Devstral Medium wins:

• Classification: 4 vs 3. Devstral Medium ties for 1st of 53 models; Gemini 3.1 Flash Lite Preview ranks 31st. This is a real, meaningful win — accurate categorization and routing tasks favor Devstral Medium.

Tied:

• Long context: Both score 4, both rank 38th of 55 — identical performance at 30K+ token retrieval.
• Agentic planning: Both score 4, both rank 16th of 54 — tied on goal decomposition and failure recovery.

The pattern is clear: Gemini 3.1 Flash Lite Preview is a broadly capable model that performs at or near the top of our tested field on most dimensions. Devstral Medium's strengths are narrow, with classification as its only outright win and competitive scores on structured output and faithfulness — but it trails significantly on safety, reasoning quality, and agentic tool use.

Devstral Medium costs $0.40 per million input tokens and $2.00 per million output tokens. Gemini 3.1 Flash Lite Preview costs $0.25 per million input tokens and $1.50 per million output tokens — 37.5% cheaper on input and 25% cheaper on output.

At 1M output tokens/month, Devstral Medium costs $2.00 vs $1.50 for Gemini 3.1 Flash Lite Preview — a $0.50 difference that's negligible. At 10M output tokens/month, that gap widens to $5.00 ($20.00 vs $15.00). At 100M output tokens/month — a realistic scale for a high-volume API integration or consumer product — Devstral Medium costs $200.00 vs $150.00 for Gemini 3.1 Flash Lite Preview, saving $50.00/month by choosing the Google model.

For developers running inference at scale, the lower cost of Gemini 3.1 Flash Lite Preview compounds meaningfully. Combined with its stronger benchmark performance, there's no cost-quality tradeoff to make here — Gemini 3.1 Flash Lite Preview is both cheaper and higher-scoring across most tasks. The cost difference matters most to teams processing tens of millions of tokens monthly; at low volumes, the gap is minor.

Choose Gemini 3.1 Flash Lite Preview if you need a general-purpose AI for production workloads: content pipelines, customer-facing assistants, multilingual applications, or any use case requiring reliable safety calibration (5/5 in our testing vs 1/5 for Devstral Medium). It's also the right call for agentic and tool-calling workflows, where it scores 4 vs Devstral Medium's 3 (ranked 47th of 54 models). The 1M-token context window gives it a substantial edge for long-document processing. At lower cost across both input and output, it's the default recommendation for nearly every workload.

Choose Devstral Medium if classification accuracy is your primary requirement — it ties for 1st of 53 models on categorization and routing in our testing, and Gemini 3.1 Flash Lite Preview scores notably lower here (3/5, ranked 31st). It also accepts only text input, which may simplify integration if multimodal capabilities would go unused and Mistral's ecosystem is already part of your stack. Note that Devstral Medium has not yet received internal benchmark scores on our full suite beyond the 12 tests shown — the description positions it as a code generation and agentic reasoning model, but our current scores do not include external coding benchmarks (e.g., SWE-bench) for either model in this comparison.