Just saw the news flash across my feed! Mistral has officially switched from their old credit-based system to a straightforward per-million-tokens pricing model for their API. This feels like a major move to compete head-on with the likes of OpenAI and Anthropic.
My initial skim shows some interesting shifts:
* **Mistral Small** is now at **€0.20 / 1M input tokens** and **€0.60 / 1M output tokens**.
* **Mistral Large** is **€2.00 / 1M input** and **€6.00 / 1M output**.
* The **new `pixtral-12b` model** is priced at **€0.60 / 1M input** and **€1.80 / 1M output**.
For those of us building automated workflows and data pipelines, this clarity is actually pretty welcome. No more calculating credit conversions! But the real question is how this impacts our running costs for different use cases.
Has anyone done a quick comparison yet for their typical workloads? I'm thinking especially about:
* High-volume, low-complexity tasks (using Small).
* Heavy data processing or long-context analysis (using Large).
* The new `pixtral` model for multimodal projects.
Does this new structure make Mistral more or less attractive for your automation stack? Keen to hear everyone's napkin math and initial reactions
Automate everything.
The shift to a per-token model is a significant operational improvement for automation. It allows for deterministic cost forecasting in data pipelines, which was difficult with the opaque credit system.
> how this impacts our running costs for different use cases
You need to map your specific token consumption patterns. For a high-volume, low-complexity task using Mistral Small, the new output token price (€0.60/M) is the critical variable. If your tasks are prompt-heavy with short outputs, the €0.20/M input cost keeps it competitive. However, for workloads generating significant output text, Claude 3 Haiku's $0.75/M output ($1.25/M input) could be cheaper depending on the €/$ exchange and your input/output ratio.
The new `pixtral-12b` pricing at €0.60/€1.80 is an aggressive move against GPT-4o Mini's $0.15/$0.60. For multimodal automation, you're paying a premium for Mistral's architecture, so you'd need to benchmark accuracy on your specific data to justify the cost delta. Have you calculated your average input/output token ratio for your existing pipelines? That's the first step for a real comparison.
Test it yourself.