Mistral Large 4 (Le Chonk) for marketing: the reasoning setting decides the bill

Checked 8 October 2026. Mistral Large 4 is in public preview, so prices, limits and licence terms may change.

The answer in brief

Mistral Large 4, nicknamed “Le Chonk”, is Mistral AI’s new flagship model, in public preview since 6 October 2026. It reads text and images; Mistral lists no image generation, and VentureBeat reports that its output is text only. You can use it now through Mistral’s API, and Mistral says downloadable weights will follow by the end of October.

• Cheap per token, not always per task. At its list price of $1.36 input and $4.18 output per million tokens, its output costs about a fifth of Claude Opus 5.5’s. Mistral currently shows a “Sale price” that halves both rates. But the model writes at length, so on Artificial Analysis’s tests a finished task costs more than on some cheaper closed models.
• The reasoning setting decides your bill. In our own tests on 8 October 2026, eight short marketing jobs cost 35 to 38 times as much at OpenRouter’s default (reasoning on) as with reasoning off, and took 20 to 29 minutes instead of about one.
• It adds selling points you did not give it. In 25 of 40 product descriptions it invented details such as “handcrafted” or “sustainable”. One sentence telling it to use only the listed facts stopped that in all 40.
• Behind the closed leaders on independent tests. Artificial Analysis scores it 38 on its Intelligence Index, against 58 for Claude Opus 5.5 and 53 for GPT-6 Astra. Cybersecurity is its strongest area.
• More say over where prompts are processed. Mistral runs its own European deployment, and open weights are promised this month, although the licence is not yet published.
• A sensible localisation trial. Its training data covered every official EU language.

What to do now: try it on a small set of real, non-confidential tasks; measure cost per approved output rather than token price; check your Mistral data settings before using client material; and wait for the licence before planning anything around self-hosting.

What is Mistral Large 4, and why is it called Le Chonk?

Mistral Large 4 (ML4) is a large language model from Mistral AI, the Paris-based AI company. Mistral’s announcement introduces it as “Unofficially ML4, very officially: le Chonk.” Mistral’s post on X on 6 October 2026 opened with “Meet Mistral Large 4, aka Le Chonk.”

The name is a size joke. “Chonk” is internet slang for a very large cat. In June 2026, Business Insider reported on a viral gag about “Le Chaton Fat”, a fictional Mistral model with “30T+ params” and “1000 meows per second”. Mistral’s CEO Arthur Mensch joined in on X: “It’s actually le gros chaton”. VentureBeat reports that Mistral executives said Le Chonk deliberately nods to that community. Mistral’s announcement gives only the nickname.

Mistral’s own pitch is that the model is “competitive with the strongest open-source models globally, while significantly outperforming any open-weight model developed in the US or Europe”. Independent tests so far put it well below the top closed models, as the sections below show.

Mistral’s announcement now describes it as “a 1 trillion-parameter natively multimodal model with 52 billion active parameters” (it said 49 billion when we first checked on 7 October), matching the 52B on its model page. Mistral’s Hugging Face page explains the gap: “49 billion active parameters per token (52 billion including embeddings and output layers)”. Its model page lists 1.05T total parameters and a 1.6B vision encoder, in a “granular Mixture-of-Experts architecture”. Only a fraction of the model runs for each token, which is how a model this large can be priced like this.

What are Mistral Large 4’s key specs?

What does Mistral Large 4 cost?

At list price, $1.36 per million input tokens and $4.18 per million output tokens. Mistral’s API pricing page and model page currently show a “Sale price” of half that, $0.68 and $2.09, next to the struck-through list price, without saying how long it lasts. In its launch note on 6 October, Artificial Analysis said that “For the first two weeks, Mistral Large 4 Preview will be served at a 50% launch discount”. By our arithmetic that ends around 20 October 2026. Budget on the list price for anything beyond a short trial.

Against other models (prices checked 8 October 2026):

• Mistral Large 3, the previous model, is listed at $0.50 input and $1.50 output on the same pricing page.
• Smaller Mistral models cost less still. The same page lists Mistral Small 4 at $0.15 input and $0.60 output. For high-volume, simple jobs such as short product titles or tagging, check whether a smaller model is good enough before defaulting to Large 4. Mistral Medium 3.5 sits at $1.50 input and $7.50 output, dearer than Large 4 even at list price, so it is not the budget step-down.
• Claude Opus 5.5, Anthropic’s newest Opus model, is $4 input and $20 output per million tokens on Anthropic’s pricing page; we cover it in our Claude Opus 5.5 guide. The earlier Claude Opus 5, which Mistral uses as the benchmark in one of its human evaluations, is $5 and $25.
• Batch and regional pricing. Mistral’s pricing FAQ on mistral.ai/pricing says batch processing “reduces the price by 50%”. Its regional inference documentation says EU or US regional endpoints are billed “at 1.1× standard list pricing (a 10% upcharge)”.

These are US list prices. Price per token is also a rough way to compare vendors, because each model splits text into tokens differently. Anthropic’s pricing page notes that its newer tokenizer “produces approximately 30% more tokens for the same text” than its previous one. Cost per finished job is the fairer test.

The headline rate understates real spend. Artificial Analysis calls the preview “very verbose”: it generated 200 million tokens across its test suite, against a median of 81 million. Mistral describes ML4 as a “hybrid instruct-and-reasoning” model, so some of those extra tokens are likely reasoning before the answer. A model that writes more costs more per finished task than its token price suggests. Artificial Analysis puts the cost per task at $1.13 at list price, or $0.57 with the discount according to its launch note. The same note called that “Over 4x the Cost per Task of similar-intelligence open weights models”. We made the same point in our Grok 4.6 review.

You can rein this in. Mistral’s reasoning documentation says Large 4 supports a reasoning_effort setting: “high” adds a full thinking step “at the cost of increased token usage”, while with “none” the model “thinks minimally”. Check what reasoning setting your router sends by default; our own test below shows what “high” costs. For routine jobs such as product descriptions, tagging or first-draft copy, test both settings and compare quality against cost per approved output. Keep “high” for analysis and multi-step work.

How good is Mistral Large 4?

Credible among open-weight models and good value, but independent scores still sit well short of the leading closed models, and most of the marketing-relevant figures come from Mistral itself.

What Mistral claims. These are the figures most relevant to marketing work, all from Mistral’s announcement:

• Business workflows: 59.9% on AutomationBench, described as “657 business workflows across apps like Gmail, Google Sheets, Slack, and Salesforce”. Mistral’s text compares it with three named rivals, all of which it beats; its own chart also includes Z.ai’s GLM-5.3, which scores higher at 62.2.
• Office deliverables: 1,393 Elo (a head-to-head rating) on AA-Briefcase, which “evaluates long-horizon knowledge work” such as spreadsheets, slides and PDFs.
• Human-rated quality: in a blind evaluation run with Surge AI, annotators rated coding outputs on a 1 to 5 scale. ML4 Preview scored 3.74, second of five models and “behind only Claude Opus 5 (4.22)”.
• Images: Mistral calls ML4 “a step change in the ability of our models to understand images” and reports 42% on the Dense 200 visual grounding test against 41% for GPT-6 Astra. Its own chart shows 42.0 against 41.5, so treat it as a tie.
• Prompt injection: ML4 “resists 93.3% of attacks” on Lakera’s B3 AI Security Benchmark.

Mistral adds that the model “continues to improve rapidly as we refine it”, so the version you test this month may not be the one that ships with the weights.

What independent testers have found so far. Artificial Analysis has published results for “Mistral Large 4 Preview”:

• Intelligence Index: 38 on version 4.3.2 (checked 8 October 2026), “above average among other reasoning models in a similar price tier (median: 26)”.
• Against the leading closed models: 58 for Claude Opus 5.5 and 53 for GPT-6 Astra at their maximum settings, on the same index (checked 7 October 2026).
• Speed: about 116 tokens per second, well above the median of about 75 for reasoning models in a similar price tier (checked 9 October 2026).
• Positioning: in its launch note, Artificial Analysis called it “the most intelligent model from outside the US and China”.

Vals AI, which runs industry-style tests, ranks it 33rd of 45 models on its Vals Index (48.05%, checked 9 October 2026), against 66.97% for Claude Opus 5.5. The strongest Vals result we found is sixth of 76 on Harvey’s Legal Agent Benchmark.

Cybersecurity is Mistral’s headline bet, and the independent data broadly backs it. On the Artificial Analysis Cyber Index, ML4 scored 50 when we checked on 6 October, fifth among the models listed: level with Z.ai’s GLM-5.3-Flash on the rounded score, and behind Grok 4.7, MiMo-V2.6-Pro and GPT-6 Luna. Its 82% on the reproduce-and-patch test is the highest there. Claude Opus 5.5 and GPT-6 Astra score close to zero on that test. Mistral says that is because they refuse the task, and Artificial Analysis’s own figures agreed when we checked on 6 October, with refusal rates of 98.5% for Claude Opus 5.5 and 100% for GPT-6 Astra. That gap reflects provider policy, not proof those models lack the ability. Mistral’s other headline cyber figure, 93% on the Cybench challenge set, is its own and has not been independently checked.

VentureBeat checked several of Mistral’s benchmark charts. It found that ML4’s coding result “looks competitive”, but “does not establish an outright coding lead across every available model-and-agent configuration”, and it could not find public sources for some competitor scores in Mistral’s visual grounding and finance charts. The official DeepSWE leaderboard (last updated 22 September 2026) does not list ML4 yet.

Our view: treat ML4 as a trial brief, not a switch memo. Mistral’s strongest claims are about cybersecurity, finance, law and engineering, and the marketing-relevant evidence (business workflows, documents, languages) is promising but mostly self-reported. A same-day test by Digitrans, a Luxembourg firm that sells legal and accounting AI tools, reported frequent grounding failures on French-language legal tasks, including a party’s argument presented as a court’s view. It is one small, secondary test, but a fair warning for any copy where qualifications matter.

How does Mistral Large 4 compare on cost and quality?

Token prices flatter Mistral Large 4 because it writes a lot. Artificial Analysis also publishes the average cost of each task in its test suite, which captures that verbosity. At the settings it tested:

Large 4’s $1.13 is at list price. With the launch discount, Artificial Analysis put it at $0.57: a little above Claude Sonnet 5.5’s $0.48, which scores higher, and still well above GPT-6.1 Sol. The test suite leans towards reasoning, coding and agentic work rather than marketing copy, so your own costs will differ. We cover both cheaper models in our Claude Sonnet 5.5 and GPT-6.1 Sol guides.

The case for Large 4 is European control and the promise of open weights, not bargain hunting.

What happened when we tested Mistral Large 4 ourselves?

The reasoning setting decided the bill, and a single sentence in the brief decided whether it made things up. On 8 October 2026 we ran two sets of tests through OpenRouter, with Large 4 routed only to Mistral’s own servers. Every prompt was invented, and the whole exercise cost us $0.94.

Eight everyday jobs. A product description, a meta description, three ad headlines, five email subject lines, a five-bullet report summary, a French localisation, product tagging into JSON and alt text for the price chart above, each with a word or character limit we could check automatically. We ran the set four times on Large 4 with reasoning off, three times at OpenRouter’s default (reasoning on), twice at “high”, and once on Claude Sonnet 5.5 and GPT-6.1 Sol at their lowest and default reasoning settings.

With reasoning off, Large 4 was the cheapest of the three models by a distance, even at list price, though a little slower than the other two. Left on OpenRouter’s default, it was the dearest and by far the slowest: the same eight jobs took 20 to 29 minutes and cost 35 to 38 times as much. A one-line meta description took between 38 seconds and more than six minutes, and up to 14,835 output tokens, with reasoning on; with it off, about a second and 24 to 30 tokens.

Reasoning did buy something. With it on, Large 4 met every limit. With it off, it missed the same two in almost every pass: the chart alt text ran to 184 to 188 characters against 125 every time (all eight figures read correctly), and one ad headline came in a character over 30 in three passes out of four. GPT-6.1 Sol met every limit at both settings; Claude Sonnet 5.5 missed two at its lowest setting and one at its default.

Ten product briefs. We then gave each model ten short product briefs with the facts listed (a stoneware mug made in Portugal, a lambswool throw, a soy wax candle and so on) and asked for 60 to 90 words for a UK homeware shop. A separate AI reviewer, given the descriptions shuffled and without model names, then checked all 120 against their facts. The counts are repeat runs of the same ten briefs, not separate products: for Large 4, each prompt got 40 descriptions (three runs with reasoning off and one with it on), and each rival got ten per prompt at its lowest setting.

• Plain prompt: Large 4 added at least one claim that was not in the brief to 25 of 40 descriptions, with reasoning on or off. Typical additions were “handcrafted in Portugal” (the brief said only “made in Portugal”), “hand-poured”, “quality British craftsmanship” on an oak chair, “100% premium linen” and “sustainable”. Claude Sonnet 5.5 did it once in ten; GPT-6.1 Sol not at all.
• One extra sentence: adding “Use only the product facts below. Do not add claims about where or how it is made, delivery, returns, guarantees, awards, sustainability or materials that are not listed” cut Large 4’s invented claims to 0 of 40.
• The catch: with that sentence and reasoning off, Large 4 wrote too little. Only 1 of 30 descriptions reached 60 words (median 40). With reasoning on it hit the range every time, at about 50 times the cost per description.

Open-ended copy brought out the same habit. For the fictional lamp shop’s meta descriptions and ad headlines, where the prompt gave no product facts at all, Large 4 claimed the lamps were “designed in Britain” or “Made in UK” at every setting, and with reasoning on it sometimes promised free delivery too, as did Claude Sonnet 5.5 (“free UK delivery over £50”). GPT-6.1 Sol did neither. Whatever model you use, give it the facts, tell it to use only those, and check every origin, delivery and product claim before anything goes live.

Structured outputs held up. With a strict JSON schema and reasoning off, all 20 calls (five messy product listings, four times each) returned valid JSON with the correct price, height, stock status and category, in a median of 3.7 seconds and for under half a cent in total.

These are small samples from one day, through a shared service whose queueing affects the times, so read the pattern rather than the exact figures. Large 4’s costs are what OpenRouter billed at Mistral’s current sale price; double them for list price. We did not test Mistral’s own API directly, and we could not test Mistral Small 4, which OpenRouter reported as rate-limited at Mistral throughout. On our free Mistral account, Large 4 did not appear in Studio’s Playground model list, and Vibe offered no way to choose it.

How we ran it. Every call went through OpenRouter’s API: Large 4 as mistralai/mistral-large-4-0, pinned to Mistral with fallbacks off, GPT-6.1 Sol as openai/gpt-6.1-sol and Claude Sonnet 5.5 as anthropic/claude-sonnet-5.5. “Reasoning off” means a reasoning effort of “none”, and “default” means we sent no setting. A script counted the words and characters, and costs are OpenRouter’s billed figures for each call. We checked limits and facts only. We did not judge whether the copy was good, on-brand or ready to use without editing, which is still an editor’s job, as in the test plan below.

Where could Mistral Large 4 help a marketing team?

These are suggestions based on Mistral’s confirmed features, not results from our own bake-off.

• Multilingual copy and localisation. Training data covering every official EU language makes ML4 worth trying for European product pages, ad variants and email campaigns. A fluent first draft still needs a native speaker before it goes live.
• Research and document analysis. With a large context window and document question answering among its API features, ML4 suits summarising reports, comparing competitor PDFs and pulling data from charts. Mistral says it “reasons powerfully across complex documents, charts, and natural images”.
• Image understanding, not image creation. ML4 can read product photos, screenshots, packaging and charts, for example to draft alt text or check listings for consistency. It produces text only, so it will not replace your image tools. VentureBeat reports that Guillaume Lample confirmed the text-only output.
• Agentic workflows. If you let agents into inboxes, spreadsheets or CRM data, AutomationBench and the prompt injection score matter. Prompt injection (hidden instructions in a page or email that hijack an agent) is a real marketing-automation risk. Reuters, via SRN News, reports Pierre Stock, Mistral’s vice president of science, saying the model “had tried to go beyond its testing environment, but that this was expected and the company was able to prevent it”. Give any agent the narrowest access it needs. If software will read the output (product feeds, CRM fields, tagging), use Mistral’s structured outputs feature rather than just asking for JSON in the prompt.

Is Mistral Large 4 good for privacy and data control?

It gives you more control options than most flagships. Check the settings anyway.

The announcement says “The model will be available across multiple regions worldwide”, including “a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law”. Mistral’s regional inference documentation says the EU endpoint runs on “Multiple data centers in EU and EFTA countries”. It also notes that “Available models vary by region”, and it does not list which models the EU endpoint serves, so confirm ML4 is offered on the endpoint you plan to use.

There is a trade-off. The same page says: “Function calling is the only supported regional tool. Stateful features, including Agents, Batch, and the Files API, are not available on regional endpoints.” So the 50% batch saving and an EU-only processing commitment do not combine today, and agent-style workflows need the global API, where “Mistral does not commit to a specific inference location”. Regional inference also covers only where the model runs. The same page says “account configuration, API keys, billing, access management, usage analytics, and other operational metadata may still be handled by Mistral systems outside the selected inference geography”.

Training and retention. Mistral’s help centre says that in Studio’s Free mode, “we may use your data (input and output) to train our artificial intelligence models” (Mistral Help Center). The opt-out guide calls the setting “Anonymous improvement data” and notes that “Vibe and API opt-out toggles are separate”. In our own free account on 8 October 2026, the API setting sat under Admin, API, Privacy with a different label, “Allow the use of your API calls to train Mistral’s AI models”, and it was switched on. Vibe’s settings are on a separate page. Zero data retention is another step again. Mistral’s ZDR documentation says it is “available on paid plans for supported stateless API calls” and must be requested.

Before anyone pastes client briefs or customer data into ML4, check that your organisation is on a paid plan, the training toggle is off and the region matches your data policy. UK businesses should confirm EU processing with their data protection lead. The ICO lists every EEA country, and Switzerland, as covered by UK adequacy regulations, which simplifies sending personal data to Mistral’s EU endpoint. A European host is not automatically the same as UK GDPR compliance for your use case.

Can you self-host Mistral Large 4?

Not yet, and probably not on ordinary hardware. Mistral’s announcement says the weights will come “by the end of the month”. Reuters reports, citing Mistral, a date of 27 October, as do VentureBeat and The Next Web. Mistral’s Hugging Face page still showed an expected release of 31 October on 8 October. Plan for late October, not a fixed day. Until then, Mistral says it is red-teaming the model with “cybersecurity leaders, vetted partners, and state authorities”.

The licence matters. Mistral’s model list shows Mistral Large 3 under Apache 2.0, a permissive licence. Large 4 is marked only “Open”. VentureBeat reports the weights “are expected under a custom Mistral license”. Don’t assume the same terms as Large 3 until Mistral publishes the licence.

On size, our rough arithmetic suggests 1.05 trillion parameters need roughly 1 TB of memory for the weights alone at 8-bit precision, or about half that at 4-bit, before any working memory. That is data-centre kit, not a laptop. For most marketing teams, open weights will matter indirectly: managed providers and agencies can host the model in specific regions, and you lean less on a single vendor. For a smaller open-weight model you can actually run, see our Muse Glimmer test.

Should your team switch to Mistral Large 4?

Our view, by situation:

• You mainly want lower bills. Test with reasoning off. On Artificial Analysis’s benchmark tasks, Claude Sonnet 5.5 and GPT-6.1 Sol cost less per task than Large 4 at list price, but on our short marketing jobs Large 4 with reasoning off was the cheapest of the three, while GPT-6.1 Sol was the most reliable on limits and facts. Test both on your own work.
• You already use Mistral Medium 3.5. Large 4 is the obvious test. Its output tokens cost less ($4.18 against $7.50 per million at list price), and Artificial Analysis scores it 38 against 14.
• You need European hosting or may want to self-host later. ML4 is the clearest new option on that brief, but wait for the licence and weights before committing.
• You need image or video generation. ML4 isn’t for that.
• You depend on a stable model for production workflows. It is a preview that Mistral says is still being trained. Keep it out of anything you can’t easily re-test.

We have also covered GPT-6 Astra, the other closed model Mistral contrasts with ML4 in its cybersecurity claims.

How should you test Mistral Large 4?

This is our recommended two-week approach, not a Mistral procedure.

1. Pick 20 real tasks across copy, research summaries, image description and one multilingual job.
2. Run them on ML4 and your current model, with the same prompts and context. Run ML4 twice, with reasoning effort set to “high” and to “none”, so you can see what the extra thinking buys.
3. Score blind. Have an editor rate accuracy, brand voice and usefulness without knowing which model wrote what.
4. Log the cost and tokens per approved output, not just the token price.
5. Check the facts. Note every error or invented detail.
6. Confirm data settings (paid plan, training opt-out, region) before using any client material.
7. Re-test after the weights release, because the model may change.

What should you check before publishing AI-written content?

A cheaper model makes it easier to produce more pages, which is exactly when review matters more. Search engines judge the published page, not the tool. Google’s guidance on generative AI content (last updated 1 October 2026) says: “It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing.” The same page says the review “also applies to metadata” such as titles, meta descriptions, structured data and image alt text. It warns that mass-producing pages “without adding value for users may violate Google’s spam policy on scaled content abuse” (spam policies). Bing’s Webmaster Guidelines say content generated at scale “without oversight, quality control, or editorial review” may be excluded from indexing.

Our guide to Google’s AI content guidance covers the review process, and our GEO guide explains how to make content easier for AI search to find, trust and mention.

If you’d like a blind bake-off of ML4 against your current model on real copy, or help setting up safe review processes, our AI adoption team can help. Get in touch to talk it through.