Mistral's Pixtral Large 2502 Is Now on OpenRouter, and Vision Models Keep Getting Cheaper
Mistral's large vision model Pixtral Large 2502 showed up on OpenRouter on November 20, priced at $2 per million input tokens with a 128,000-token context window.

Pixtral Large 2502 Hits the API Catalogs
Mistral's Pixtral Large 2502 appeared on OpenRouter on November 20, priced at $2 per million input tokens and $6 per million output tokens, with a 128,000-token context window. It takes text and images in and returns text, which makes it a workhorse for document reading, chart analysis, and anything where a model needs to look at a picture and explain what it sees.
One honest note before the hype: this is not a brand-new model. Pixtral Large 2502 is the February 2025 version of Mistral's large multimodal model, the successor to Pixtral Large 2411, and its arrival on OpenRouter is a distribution milestone rather than a research breakthrough. That distinction matters, because the story here is about access and price, not a new capability nobody had before.
The model is also listed through Amazon Bedrock, which puts it in front of enterprise buyers who already run their AI workloads on AWS. For a European lab trying to sell into companies that care about data residency and vendor diversity, showing up in the catalogs people already use is half the battle.
What Pixtral Large 2502 Is Good At
Vision models earn their keep on unglamorous work. Reading a scanned invoice, pulling numbers out of a chart, checking a photo against a written description, or summarizing a 40-page slide deck: these are the jobs where a model that can see pays for itself. Pixtral Large was built for exactly that kind of multimodal grind, and the 128K context window means it can hold a long document and its images in one pass instead of chunking everything into fragments.
The pricing tells you who Mistral is aiming at. At $2 per million input tokens, Pixtral Large 2502 costs a fraction of what the top frontier vision models charge, and it undercuts several mid-tier options too. For teams processing thousands of documents a day, that gap adds up fast. The trade-off is raw capability: Pixtral Large is a strong model, but it is not trying to beat Gemini 4 Neon or GPT-6.2 Astra on the hardest reasoning benchmarks. It is trying to be good enough at a price that makes large-scale document work affordable.
That is a sensible place to compete. Most companies do not need the smartest model on earth to read their paperwork. They need one that is accurate, fast, and cheap enough to run on everything.
The Vision Model Price War Is Real
Pixtral Large 2502 arrives in the middle of a crowded week. Google shipped Gemini 4 Neon on November 18, OpenAI's GPT-6.2 family landed the day before, and open-weight challengers keep appearing on the same API catalogs. The pattern is consistent: every lab now ships tiers, and the middle tier is where the volume lives.
Mistral's edge is its positioning. It is the European challenger with open-weight roots, and it has spent the year building a full stack, from small on-device models to large API models, instead of betting everything on one flagship. Pixtral Large 2502 fits that strategy: a capable, affordable vision model that slots into existing pipelines without drama.
The risk is obscurity. A February 2025 model arriving on a major API catalog in November 2026 does not generate the launch-day buzz of a fresh release, and Mistral has to work harder to explain why a model that is nearly two years old deserves a look. The honest answer is the price and the context window, and for the document-heavy workloads this model targets, that answer is probably enough.
The Bottom Line on Pixtral Large 2502
If you process documents, charts, or images at scale, Pixtral Large 2502 is worth a test run. The $2 per million input token price and 128K context make it one of the cheaper ways to put a capable vision model on real work, and its availability on OpenRouter and Bedrock means you can try it without reworking your stack. It will not win every benchmark, and Mistral is not pretending otherwise. But for the everyday job of making machines read what humans used to read, cheap and good enough beats expensive and brilliant most days of the week.


