Open, European, and cost-effective: Why we chose Mistral models as our default
Most conversations about large language models (LLMs) revolve around the same two or three big US providers. But there’s a strong European alternative that deserves more attention: Mistral AI. At Getsper, Mistral models are the default models powering our workflow engine, and we’ve spent months running real, everyday automation through them. Here’s what that’s actually been like.
Mistral AI is a French company, and its models are among the most credible European alternatives to the well-known US-based AI labs. That matters for three reasons: transparency, several models are released open-weight under permissive licenses like Apache 2.0, so the weights are publicly available for anyone to inspect, verify, or self-host, instead of trusting a black box; supporting the European AI ecosystem, backing a startup in a field dominated by a handful of US companies; and strong performance for the price, which the numbers below back up.
The Three Models used, tested and run in Production
Mistral Large 3 (December 2, 2025) is Mistral’s state-of-the-art, general-purpose model, open-weight under Apache 2.0, and multimodal. It uses a Mixture-of-Experts architecture, packing 675 billion total parameters while only around 41 billion are “active” per request.
Context: 256k tokens · Pricing: €0.44 / €1.3 per million input/output tokens · License: Apache 2.0
Mistral Medium 3.5 (April 28, 2026) is a frontier-class multimodal model optimized specifically for agentic and coding use cases, which matters a lot for Getsper, since our workflow engine takes actions, calls tools, and completes multi-step tasks autonomously. Open weights under a Modified MIT license.
Context: 256k tokens · Pricing: €1.25 / €6.4 per million input/output tokens · License: Modified MIT
Mistral Small 4 (March 16, 2026) is the most budget-friendly of the three and impressively capable for its size, a “hybrid” model unifying instruction-following, reasoning, and coding into one, with 119 billion total parameters but only 6.5 billion active at a time.
Context: 256k tokens · Pricing: €0.12 / €0.5 per million input/output tokens · License: Apache 2.0
Everything above describes these models in their base form, with no fine-tuning or customization, we’re evaluating them exactly as Mistral ships them.
Our Real-World Experience
Specs are one thing; what matters is how a model performs wired into a real product, handling real tasks, day after day.
Tool calling, the backbone of any agentic system, has been reliable and consistent across all three models, a model that fumbles tool calls breaks entire workflows, and we haven’t seen that here. Test generation has consistently produced solid, logically structured output without requiring heavy manual correction. Database management tasks, which demand structured and precise handling, have run smoothly. And task management, planning, sequencing, and tracking multi-step work, has been strong enough that we made these models our default across the platform.
This isn’t a “looks good in a demo” situation. It’s day-to-day production usage across a range of task types, and the models have held up.
How Mistral Compares on Cost
Put Mistral’s per-token pricing next to other major LLM providers and the gap is significant — Mistral is considerably cheaper across the board, especially Small 4 at €0.12 / €0.5 per million tokens. For a platform running large volumes of agentic workflows, where every tool call and reasoning step consumes tokens, that difference compounds fast and changes what’s economically viable to automate in the first place.
The Verdict
After running Small 4, Medium 3.5, and Large 3 as the default models inside Getsper’s workflow engine, our takeaway is straightforward: this is a strong, well-rounded package, solid real-world performance on tool calling, test generation, database work, and task orchestration, combined with open-weight transparency, support for a European AI company, and pricing that’s hard to beat. No model wins on every dimension, but for teams looking for genuine value, Mistral’s current lineup is well worth a serious look. It’s exactly why we made it the default at Getsper.
Content research by us. Text created by AI