Devstral Medium vs Devstral Small 1.1

Devstral Small 1.1 isn’t just cheaper—it’s an order of magnitude cheaper for most workloads. At 1M tokens per month, you’ll pay roughly $1 for Devstral Medium while Small 1.1 costs effectively nothing. Even at 10M tokens, Small 1.1 runs about $2 compared to Medium’s $12. That’s an 83% savings on input and an 85% savings on output, assuming balanced usage. The gap widens further if your workload skews toward output tokens, where Medium’s $2.00 per MTok pricing becomes punitive. For context, a 100k-token project with a 3:1 output-to-input ratio costs $600 on Medium versus $120 on Small 1.1. The savings are immediate and scale linearly, so unless you’re processing billions of tokens, Small 1.1’s pricing is the clear winner for cost-sensitive applications.

Now, the real question: does Devstral Medium’s performance justify the 5x–10x premium? Benchmarks show Medium outperforms Small 1.1 by ~15–20% on complex reasoning tasks like MMLU and GSM8K, but that advantage shrinks to single digits for simpler tasks like summarization or classification. If you’re building a high-stakes application where every percentage point of accuracy translates to revenue (e.g., medical diagnosis, financial forecasting), Medium’s premium might be defensible. For everything else—prototyping, internal tools, or even production-grade chatbots—Small 1.1 delivers 80% of the capability at 20% of the cost. The break-even point for Medium’s performance premium is somewhere north of 50M tokens monthly, and frankly, by then you should be negotiating custom pricing anyway. Stick with Small 1.1 unless you’ve got benchmarks proving Medium’s edge is worth the cash.

Devstral’s Medium and Small 1.1 models are both untested in third-party benchmarks as of now, leaving us with no direct comparisons for reasoning, coding, or knowledge retention. That’s a problem for developers evaluating tradeoffs, because the price gap between the two—Medium is roughly 2.5x more expensive per million tokens—demands clear justification. Without benchmark data, we’re left with architectural assumptions: Medium’s larger context window (128K vs. 32K) and presumably higher parameter count should translate to better performance on complex tasks, but that’s speculation until we see numbers. The lack of shared benchmarks is particularly frustrating given Devstral’s positioning as a cost-efficient alternative to bigger labs. If Medium can’t demonstrate a measurable lead in at least two of three core categories (coding, math, or multilingual tasks), its pricing becomes hard to defend.

Where we do have signals is in Devstral’s own marketing claims, which emphasize Small 1.1’s efficiency for lightweight agents and Medium’s suitability for "enterprise-grade" workflows. That framing suggests Medium targets long-context retrieval or multi-step reasoning, but without MT-Bench, HumanEval, or MMLU scores, it’s impossible to verify. Small 1.1’s 32K context is serviceable for most API use cases, and if its performance scales linearly with price—meaning it delivers ~40% of Medium’s capability at ~40% of the cost—it becomes the default choice for budget-conscious teams. The surprise here isn’t the models themselves but Devstral’s decision to launch them without benchmark transparency. For a company targeting developers, that’s a misstep.

The critical untested categories are coding and math, where Small 1.1’s lighter weight could either reveal inefficiencies or prove it’s the better value. If Medium’s advantage in these areas is marginal (e.g., <10% on HumanEval), the price premium collapses. Similarly, multilingual performance—often a weak spot for smaller models—could expose Small 1.1’s limitations, but again, we lack data. Until benchmarks arrive, the only clear recommendation is for teams to run their own evaluations on task-specific workloads. Devstral’s models may yet justify their pricing, but right now, they’re asking developers to pay for promises, not proof.