Mistral Large 4 Review: Official Pricing $1.36/$4.18, 1M Context, AA Intelligence Index 38, Real-World Coding and Document Performance

The bottom line:Mistral Large 4 is the flagship model Mistral released as a public preview on October 6, 2026. The API model name is mistral-large-4, with official pricing of $1.36 per million input tokens, $0.14 for cached input, and $4.18 for output, and a 1M context window.

Its strengths and weaknesses are both clear. Official and third-party data both show it is strong incybersecurity, finance, and document understanding, but its composite Artificial Analysis Intelligence Index score is only 38. Itthinks a lot and outputs slowly, and coding tests from several developers also put it behind models like Qwen3.8 Max, Kimi K3, and DeepSeek V4.1 Flash.

Figures in this post are current as of 2026-10-22, based on Mistral's official launch blog,model docs,pricing page, andchangelog, plus Artificial Analysis,OpenRouter and other third-party pages. Leaderboard scores change, so defer to the original pages.

What Is Mistral Large 4

Mistral's official description:

Launched on October 6, 2026 as a Public Preview release, nicknamed Le Chonk, and the largest and strongest model Mistral has built to date
The docs state it is amixture-of-experts (MoE) architecture with 1.05T total parameters and 52B active parameters, plus a 1.6B vision encoder, with native image input support
One model covers instruction following, reasoning, and agent tasks, with reasoning depth controlled by the reasoning_effort parameter
Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, with training data covering more than 160 languages
Official focus areas: cybersecurity, coding agents, professional work such as finance and law, and visual grounding
Weights planned to open at the end of October. As of 2026-10-22, Mistral's Hugging Face page does not yet have Large 4 weights, so for now it is only available through the API

The parameter counts are not fully consistent: the docs say 1.05T total / 52B active, the launch blog says "1 trillion / 52 billion active", and Mistral's official X post says "49B active". This post follows the docs.

Official Pricing

Mistral pricing page, per million tokens, in US dollars:

mistral-large-4	Price
Input	$1.36
Cached input	$0.14
Output	$4.18

A few notes:

The official changelog says50% off for the first two weeks(starting October 6). Some aggregator pages still show a discounted $0.68 / $2.09; go by the official pricing page's $1.36 / $4.18
Cached input costs about 1/10 of the standard rate, which saves a lot for long system prompts and workloads that repeatedly read the same documents
The pricing page also offers euro pricing plus Batch and Priority tiers; check the official pricing page for details
Compared with Mistral's other models: Mistral Large 3 is $0.5 / $1.5, and Mistral Medium 3.5 is $1.5 / $7.5. Large 4's output price is even lower than Medium 3.5's

The unit price looks low, but itthinks a lot, so actual cost depends on output token usage. See the benchmarks section below.

Once it is listed on AI Code With, we will add our pricing here.

Context and Capabilities

Item	Details
Model name	mistral-large-4(alias in the docs: mistral-large-4-0)
Base URL	https://api.mistral.ai/v1
Context	1M (OpenRouter lists 1,048,576 tokens)
Max output	262,144 tokens (from OpenRouter and models.dev; not stated separately in the official docs)
Input	Text, images
Output	Text
Reasoning control	reasoning_effort:high returns the full reasoning content,none does minimal thinking and returns no reasoning content
API features	Structured outputs, function calling, document Q&A, Prefix, Batch, Agents and Conversations, built-in tools
Weights	Planned for the end of October; API only for now

The context figures are not fully consistent: Mistral's docs and OpenRouter say 1M, the Artificial Analysis page says 524k, and channels such as Vercel and opencode also list 524,288. If you need to fit in more than 500k tokens, first confirm how long a context your channel actually supports.

Two points on reasoning, from Mistral's official reasoning docs:

reasoning_effort set to high , the returned message.content is a list of chunks (thinking chunks plus text chunks), not a plain string, so your parsing code needs to handle that
In multi-turn conversations, put the full assistant message (including thinking chunks) back into the history unchanged. The docs explicitly say that removing thinking chunks noticeably degrades output quality

Benchmarks and Hands-On Tests

Official Data

All of the following are published by Mistral itself, with no third-party verification:

Category	Benchmark	Mistral Large 4
Coding	DeepSWE v1.1	61.7%
Coding	SWE-Atlas-QnA	59.4%
Coding	Terminal-Bench 4	28.3%
Coding	Coding Agent Index	49.8% (Mistral says this is higher than DeepSeek V4 Pro 0813 and Qwen3.8 Max)
Agent	AutomationBench	59.9%
Knowledge work	AA-Briefcase	1,393 Elo
Cybersecurity	CyberGym-E2E	82%
Cybersecurity	Cybench	93%
Visual grounding	Dense 200	42% (GPT-6 Astra 41%)
Safety	Lakera B3 attack resistance	93.3%

Mistral also gave two comparisons:

In a blind coding-quality evaluation by Surge AI (scored 1-5), the Large 4 preview scored 3.74, ranking 2nd among 5 models, above Kimi K3 (3.59), GLM-5.3 (3.60), and GLM-5.2 (3.40), and below Claude Opus 5 (4.22)
On cybersecurity, it ranks in the global top five on the Artificial Analysis Cyber Index. Mistral specifically notes that Claude Opus 5.5 and GPT-6 Astra score close to 0 on tests like "reproduce the vulnerability, then patch it", because they refuse to run them outright

Third-Party Data

Artificial Analysis(tested on the Preview reasoning version):

Intelligence Index 38, versus a same-price median of 27. For comparison, on the same index Claude Opus 5.5 scores 57.6, GPT-6.1 Sol scores 51.8, and Gemini 4 Argon (High) scores 52.6
Running the full index took 200M output tokens, versus a same-price median of 87M. AA's verdict: "very verbose"
Output speed is 70.9 tokens/sec, below the same-price median of 79.2; time to first token is 1.92 seconds
Average cost per task is $1.13
Breakdown: long-context reasoning AA-LCR 81.3%, Humanity's Last Exam 35.0%, GDPval-AA 46.1%, SciCode 54.2%, Terminal-Bench 4.0 26.8%

Vals AI(republished by OpenRouter):

Vals Index 48.7%, Vals Finance Index 56.9%, Finance Agent v2.1 61.4%
Coding: Terminal-Bench 4.0 27.8%, SWE-bench Multimodal 33.1%, Vibe Code Bench v1.1 78.4%

OpenRouter: P50 throughput on the best provider is about 55 tokens/sec.

Hands-on results from developers

Surge AI's sudo L7(a coding benchmark that tests senior-engineer judgment): 13.3%, ahead of Muse Spark 1.3 (12.2%) and Grok 4.7 (10.6%)
Paweł Huryn's bug-finding test: 105 bugs planted across two real repos. Large 4 (high setting, which is also its highest) averaged 15 points over three runs, versus 25.7 for Qwen3.8 Max, 23 for Kimi K3, 21.7 for DeepSeek V4.1 Flash, and 19 for GLM-5.3
Model Tested: 99.7 on document reading (out of 100), higher than every local model they have tested. But it solved only 18 of 30 coding problems,and on the other 12 it used up its thinking budget without producing an answer. Lowering the reasoning level didn't help. All 18 solved problems were correct
TheGitAI: after testing, they decided not to plug it into their own coding agent, because its score was below GLM 5.3 Flash and DeepSeek V4.1, and it was the slowest model they tested
Silicon AI Software Factory(enterprise .NET/Vue project): ranked 14th on bug fixing and 28th on implementing user stories

The hands-on results agree:it does well on document, finance, and security tasks, writes code worse than comparable Chinese models, and tends to think for too long.

Strengths, weaknesses, and where it fits

Strengths:

It will do cybersecurity work: tasks like reproducing vulnerabilities, writing patches, and analyzing malware, which many closed-source models refuse. Large 4's official scores and positioning lean this way
Strong on documents and finance: Model Tested's document reading score is near perfect, and its Vals finance scores are solid too
Low unit price: output is $4.18, even lower than Mistral Medium 3.5, and cached input is only $0.14
1M context + image input, with official visual grounding numbers slightly above GPT-6 Astra
European deployment, open weights planned: appealing to teams with data sovereignty or self-hosting requirements

Weaknesses:

Thinks too much: AA measured output token volume at more than double the same-price median, and a cheap unit price doesn't mean a cheap total bill. One user comparing results under Mistral's official post found that on the same revenue-forecasting coding task, Large 4 cost about $14 while GPT-6.1 Sol cost about $1
Slow: output speed measured by both AA and OpenRouter falls between 55 and 71 tokens/sec
Not a strong coding agent: multiple hands-on tests put it behind Qwen3.8 Max, Kimi K3, DeepSeek V4.1 Flash, and GLM-5.3
Still a preview: Mistral says the model is still being trained, so scores and behavior may change; context length is labeled inconsistently across providers

Good fits:

Security teams doing vulnerability research, code security audits, and detection rule writing
Reading and extracting from long documents such as finance, legal, contracts, and reports
Companies with data residency requirements that want to self-host later
Poor fits: latency-sensitive online scenarios, or serving as the primary coding model in Claude Code / Codex

How to access it

AI Code With doesn't list Mistral Large 4 yet. Once it does, we'll add our model id and configuration steps here.

For now you can use Mistral's official API directly (create a key in Mistral Studio). Chat Completions example:

bash

curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-large-4",
    "messages": [{"role": "user", "content": "你的问题"}],
    "reasoning_effort": "high"
  }'

Python SDK:

python

import os
from mistralai import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
resp = client.chat.complete(
    model="mistral-large-4",
    messages=[{"role": "user", "content": "你的问题"}],
    reasoning_effort="high",
)
print(resp.choices[0].message.content)

The model name on OpenRouter is mistralai/mistral-large-4-0, using an OpenAI-compatible interface.

A few usage tips:

For simple tasks, try reasoning_effort set to nonefirst, which saves tokens and is much faster; turn on high
for hard problems. Give max_tokens plenty of headroom and set a request timeout, so a long stretch of thinking doesn't end with nothing returned
In multi-turn conversations, put the full assistant message back into the history, not just the body text
Before using it in production, compare cost and latency against GLM-5.3 and DeepSeek V4.1 Flash on your own tasks

For the general setup of third-party APIs, see "How to Connect Claude Code to a Third-Party API" and "Codex config.toml Configuration Guide".

FAQ

Is Mistral Large 4 open source?

It's officially positioned as an open-weights model, with weights planned for release at the end of October. As of 2026-10-22 they haven't been released, so for now you can only use the API.

How much does Mistral Large 4 cost?

Official pricing per million tokens is $1.36 for input, $0.14 for cached input, and $4.18 for output. The 50% off launch promotion for the first two weeks has ended.

Is Mistral Large 4 good for coding?

It works, but it isn't its strongest area. Its official DeepSWE v1.1 score is 61.7%, it generally trails Qwen3.8 Max, Kimi K3, and DeepSeek V4.1 Flash in third-party tests, and it tends to think too long. For primary coding work, consider other models first.

What is Le Chonk?

It's a nickname for Mistral Large 4, which Mistral's own official blog uses.

What Is Mistral Large 4? Official Pricing, 1M Context, Benchmarks, and Hands-On Tests: Strong at Cybersecurity and Documents, Slower at Coding