# Pixtral Large 2502: Mistral Model Specs and Pricing

By Ossian Kettunen · 2026-11-20 · Source: https://www.activepieces.com/blog/pixtral-large-2502-mistral-model-specs-and-pricing

---
<aside class="tldr"><p class="tldr-label">Summary</p><p>Pixtral Large 25.02 offers no advantage for automation tasks, with worse accuracy and 2.4 times higher costs than available alternatives.</p><ul><li>Pixtral Large succeeded on 2 of 3 tasks; GPT-5.4-mini succeeded on all three.</li><li>Pixtral Large cost $0.000361 per task, 2.4 times more than GPT-5.4-mini. -</li></ul></aside>

For teams wiring models into flows with tools like [Activepieces](https://www.activepieces.com), the first job is pinning down what this model is. Pixtral Large 25.02 is an open-weights multimodal model from [Mistral](https://mistral.ai/news/pixtral-large) AI, built on Mistral Large 2.

## Pixtral Large 25.02 is an older Mistral multimodal model

### What Mistral released and when

Here is what you'd be swapping in, and the point to take is that this is a large model, not a small one. Mistral's page, dated November 18, 2024, gives these facts.

- Maker: Mistral AI
- What it is: open-weights multimodal model, built on Mistral Large 2
- Size: a **123B-parameter decoder plus a 1B vision encoder**
- Page date and check date: 2026-11-20

Almost all of the weight sits in the text decoder. The vision encoder's a small add-on, so this is Mistral Large 2 given eyes rather than a new family.

### How to get it, per Mistral's page

Because the page lists per-token prices, hosted use is metered. Pixtral Large costs [$2 to $6 per million tokens](https://docs.mistral.ai/models/pixtral-large-24-11.md), the same range as Mistral Large 2, so adding image input doesn't raise the rate.

At $0.15 per million tokens at both ends, Pixtral 12B was the cheap option at the time; Ministral 3 8B is Mistral's current efficient vision option.

That made it roughly **13 times cheaper at the low end** and 40 times at the high end than Pixtral Large.

You could host it yourself, since the weights are open, but a 123B decoder implies serious GPU hardware, and that cost is yours to count.

## Why a vision-capable model matters to automation teams

A vision model lets a workflow act on what a file shows, not just on text someone typed into a field. That decides which steps you can automate at all.

Most automations break at the same place: **the data arrives as pixels**, and a person has to retype it before the next app can use it.

Pixtral Large 25.02 is an older Mistral multimodal model, so it reads images as input and returns text your workflow can map into fields; Mistral Large 4 is the current multimodal flagship.

Where this matters most is four inputs, each paired with the app that receives the result:

- An invoice image, read for vendor, date and totals, then sent to the accounting system.
- A ticket screenshot, read for the error message and product, then sent to the helpdesk.
- A scanned form, read field by field, then sent to a spreadsheet row.
- An attachment on an incoming email, read for what it's, then sent to the CRM record.

![A scanned invoice image on the left with vendor, date and totals areas marked by highlighted boxes, and an arrow leading to…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/78f5a6f3-18ea-4355-a819-9909fee2ef7f/pixtral-large-2502-mistral-model-specs-and-prici-aae80dab.webp)

In each case the model only does the reading. The field mapping on the receiving side still has to be built, and it's where these workflows usually fail: a total read into the wrong column is worse than no automation.

So the useful test of a swap is not whether the model can describe an image, but whether its output lands in the right field, consistently, on your own documents.

## What Mistral says is new in Pixtral Large 25.02

Mistral presents Pixtral Large as an open-weights multimodal model built on Mistral Large 2, and its stated figures cover scale and one benchmark, not whether a field gets filled correctly.

The table below lists what the Mistral AI announcement page, dated November 18, 2024, claims, so you can see what's stated and what's not.

### Capabilities and figures Mistral states

| Fact | Value | Source | Checked |
|---|---|---|---|
| Parameters | 124B | [Mistral AI](https://mistral.ai/news/pixtral-large) | 2026-11-20 |
| Context window | 128K | [Mistral AI](https://mistral.ai/news/pixtral-large) | 2026-11-20 |
| MathVista | 69.4%, vs GPT-4o 65.4% and Gemini 1.5 Pro 67.8% | [Mistral AI](https://mistral.ai/news/pixtral-large) | 2026-11-20 |
| Licence | Open weights | [Mistral AI](https://mistral.ai/news/pixtral-large) | 2026-11-20 |

![A table of columns on an accounting screen where a total value from a scanned invoice has been placed in the wrong column…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/3bb82ae6-c8bc-43db-a050-cad1d9e6d520/pixtral-large-2502-mistral-model-specs-and-prici-9aee1686.webp)

At 124B parameters, this is a heavy model to self-host, so the open weights help only if you have the hardware.

With a 128K window, long, multi-page documents can go in one call, but the window says nothing about accuracy deep into it.

MathVista tests mathematical reasoning over images, and the two models it beats are older ones, so it's a weak proxy for reading invoices or forms.

### Pricing and access terms on Mistral's pages

Mistral lists per-token pricing for Pixtral Large alongside its other API models, Pixtral 12B and Mistral Large 2. Those rates are not confirmed here, so none appear. Read them on Mistral's pricing page before you cost out a swap.

![Pixtral Large API pricing against other Mistral models](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/db4631ce-3b30-4ef6-af10-bd3edaacc8e7/pixtral-large-2502-mistral-model-specs-and-prici-27ce861c.svg "Source: Mistral docs")

A larger model's likely to cost more per call than the 12B model, and at volume that gap decides whether the swap pays for itself.

## What Activepieces has already shipped for Pixtral Large

[Activepieces](https://www.activepieces.com/pieces/mistral-ai) has shipped a general Mistral integration, and nothing in it is specific to Pixtral Large. The [Activepieces catalog entry](https://www.activepieces.com/pieces/mistral-ai) (`@activepieces/piece-mistral-ai`) describes open-weight and hosted language models for text generation, embeddings and reasoning, with 6 actions and 0 triggers.

![A modal dialog for installing a custom piece in Activepieces, with fields for package type, piece name, and version.](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/a568b71d-b5b8-4e98-bcd9-834b9227ff0a/self-host-deepseek-r1-for-private-automation-202-7d9a246c.webp)

With no triggers, the integration can't start a workflow, so it only works as a step after something else fires. Nothing in that description mentions image input, so test it before you build on it.

Whether Pixtral Large earns a step is also your decision to make, not ours.

A platform that resells its own model has already picked and priced that swap before you open the product.

 Mistral's current flagship open-weight multimodal model is Mistral Large 4 (v26.10), so Pixtral Large's an older generation. That matters for the swap: if you're choosing a vision model today, the newer one belongs in the comparison.

Check Mistral's own model page for the exact model identifier and access terms, then put that identifier into the piece's action. If the action's model field doesn't accept it, the integration's the limit, not the model.

Below is the admin screen: the Integrations page with the "Install a integration" dialog open, set to NPM Registry, with fields for the package name and version.

It is the route for adding a integration that isn't in your instance yet, and it needs the admin interface.

Once the integration's installed, the work's mapping an image field from your trigger into the Mistral action.

## How Pixtral Large performed on three tasks

On 2026-11-20, Pixtral Large 25.02 and openai/gpt-5.4-mini were run on the same three automation tasks, so the results are directly comparable. Each was checked against a known answer, which gave a pass or fail per task.

The tasks were sorting a support ticket, pulling fields from an invoice, and summarising an email thread. Both models ran through [OpenRouter](https://openrouter.ai/mistralai/pixtral-large-2502) at default settings. Three tasks is a small sample, so read the results as a first look, not a ranking.

![Task speed: Pixtral Large against GPT-5.4 mini](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/aa8e4042-c30b-4747-9844-06574bfca176/pixtral-large-2502-mistral-model-specs-and-prici-405d1981.svg "Source: Activepieces own test, 2026-11-20 (2026)")

### Pixtral Large accuracy, speed and cost per task

Two of three is what Pixtral Large got right, and GPT-5.4 mini got all three. On a sample this size that's a warning to check your own cases, not a measured error rate.

Pixtral Large had a median of **0.8 seconds per task at $0.000361 a task**, and GPT-5.4 mini had a median of 0.9 seconds at $0.000150 a task.

Per task, Pixtral Large took 0.5 seconds on ticket sorting against 0.9 for GPT-5.4 mini, so it was faster on the lightest job.

On invoice fields it took 1.2 seconds against 1.1, and on the email summary 0.8 against 0.9, meaning performance varies by task. Those gaps are a tenth of a second or less, which no workflow will notice.

Cost's the larger gap. At $0.000361 a task, Pixtral Large runs about 2.4 times the price of GPT-5.4 mini. Over 10,000 tasks that's $3.61 against $1.50.

**On this test, then, speed gives no reason to swap, and accuracy and cost both favour the model already in place.**

## How Pixtral Large changes cost and speed versus today's options

Cost points the same way as accuracy. Pixtral Large cost about 2.4 times as much per task as gpt-5.4-mini in the test, at the same speed, so the swap buys nothing on price.

The chart shows $0.000361 per task for Pixtral Large against $0.000150 for gpt-5.4-mini, with median times of 0.8 and 0.9 seconds on the bars, which means **Pixtral's cost premium far outpaces its speed advantage**. Pixtral Large's the more expensive choice, and it got fewer answers right on this test.

At the volumes automation teams run, that gap compounds with every call, and the 0.1-second edge's too small to offset it.

 Mistral's current flagship multimodal model is Mistral Large 4, so this is an older release, and any comparison should be read with that age in mind.

You can call it through Mistral's API or download the weights and host it yourself. Self-hosting's where the bill moves from tokens to hardware.

[GetDeploying](https://getdeploying.com/llms/pixtral-large) lists Pixtral at 262 GB of GPU memory at 16-bit, against 180 GB for Llama 3.2 Vision 90B, so you need roughly 80 GB more just to load it.

![GPU memory to self-host Pixtral Large and Llama Vision](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/40677259-719e-46dc-a117-66d4f30be58f/pixtral-large-2502-mistral-model-specs-and-prici-5e7eddd8.svg "Source: GetDeploying")

At 8-bit the figures are 138 GB for Pixtral and 90 GB for Llama, which puts Llama on a single 96 GB card while Pixtral needs a multi-GPU node.

At 4-bit, Pixtral needs 82 GB and Llama 56 GB. Even the smallest Pixtral setup **exceeds what one 80 GB card can hold**, and Llama fits.

## What to watch next after the Pixtral Large 25.02 release

Watch the facts that can still change, such as access terms and hosting requirements. Rerun the comparison on your own documents, because the hardware gap above means a swap's not free.

Multimodal means the model reads images alongside text, and Pixtral Large 25.02 is one. Its tag follows Mistral's year-month versioning, so treat it as one dated snapshot and not a permanent name.

Since then Mistral has listed newer models, and any of them may change which one deserves your test slot. Check the date on every claim you rely on before you quote a job around it.

Access terms, licence conditions and hosting options are exactly what a vendor rewrites quietly, so Mistral's own page's the source to trust over any summary, including this one.

To keep a swap decision honest, follow the order below: confirm the facts, then measure on your own data. Each step narrows what you still don't know.

1. Recheck Mistral's model page for access route, licence terms and hosting requirements.
2. Pick real documents from your workflow, including the ugly scans and odd layouts, so the test reflects what the automation actually sees.
3. Run both your current model and Pixtral Large against answers you already know are correct.
4. Compare how many answers were correct, how many seconds each took, and what each run cost.

A swap that wins on correct answers but loses on seconds or cost is a trade, not an upgrade. Decide which of the three your workflow can't give up before you read the results.

![A three-legged stool with legs of different lengths and a saw.](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/ebf64425-c5ee-4777-976c-3952056e405a/pixtral-large-2502-mistral-model-specs-and-prici-16c23dc0.webp)

## Related reading

- [What is Mistral Large 4? Pricing and Benchmarks](https://www.activepieces.com/blog/what-is-mistral-large-4-pricing-and-benchmarks)
- [Mistral Large 4: A Guide to the New Frontier Model](https://www.activepieces.com/blog/mistral-large-4-a-guide-to-the-new-frontier-model)
- [GPU Requirements for Self-Hosting Mistral Large via API](https://www.activepieces.com/blog/gpu-requirements-for-self-hosting-mistral-large-via-api)

## References

- [GetDeploying](https://getdeploying.com/llms/pixtral-large)
- [Activepieces own test, 2026-11-20](https://openrouter.ai/mistralai/pixtral-large-2502)
- [Mistral docs](https://docs.mistral.ai/models/pixtral-large-24-11.md)
