🔍 Read the full analysis: Mistral Large 4 Lags Behind The Current AI Frontier on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Mistral launched Mistral Large 4 as an API preview on Oct. 6, with a public release of model weights scheduled for later in October. Artificial Analysis rated it 38 on its Intelligence Index, below several leading US and Chinese models; the comparison is a dated benchmark snapshot, not a direct measure of performance on every task.
Mistral says Large 4 is a mixture-of-experts model with one trillion total parameters and 49 billion active parameters. It can process text and images. The company says it trained the model on its own infrastructure in Europe and is continuing to improve it. For now, developers can access the preview through an API; Mistral has scheduled the model weights for release later in October.
Artificial Analysis assigned the preview an Intelligence Index score of 38. In the Oct. 7 snapshot cited by ThorstenMeyerAI.com, that matches OpenAI’s GPT-6 Luna at maximum reasoning effort, sits one point below DeepSeek V4.1 Flash at maximum effort, and trails higher-scoring models including Anthropic’s Claude Opus 5.5, Google’s Gemini 4 Argon and OpenAI’s GPT-6.1 Sol.
The same comparison lists China’s Z.ai GLM-5.3 at 45 and Moonshot AI’s Kimi K3 at 44. Canada’s Cohere Command A+ scores 13, a counterexample to any blanket claim that all major competitors rank ahead of Mistral. The source cautions that developer locations do not identify where API requests are processed, and the compared reasoning settings are not evaluations under identical compute budgets.
AI FRONTIER · BENCHMARK SNAPSHOT · OCT 7, 2026
Mistral Large 4 Lags Behind The Current AI Frontier
Mistral’s new model is available as an API preview, with public weights planned for later in October. An early benchmark places it below several leading US and Chinese models—one signal for developers, not a verdict on every task.
Oct. 7 comparison
The index aggregates results across a benchmark suite. It does not directly predict success or reliability on every real world workload.
At a glance
A preview, with a release still ahead
Mistral says it trained Large 4 on its own infrastructure in Europe and continues to improve it.
API first
Announced Oct. 6, 2026. Developers can access the preview through an API; the weights were not yet publicly downloadable in the reported snapshot.
Below several leaders
Its score of 38 trails higher scoring models in the Oct. 7 comparison, while matching GPT-6 Luna at maximum reasoning effort.
A dated benchmark
Scores and model versions can change. Different reasoning settings may also involve different compute budgets.
The benchmark gap
38 sits well below the leading scores
Artificial Analysis Intelligence Index scores in the cited Oct. 7, 2026 snapshot. Bar lengths show index points, not percentages or task success rates.
The comparison includes models and reasoning settings that are not necessarily evaluated under identical compute budgets. Developer location does not establish where API requests are processed.
What the preview cannot establish
One score cannot settle a workload decision
Not a task-by-task forecast
A score of 38 does not prove Large 4 will fail a specific coding, research, or professional assignment. Teams need tests that reflect their own work.
Not a controlled reliability study
Thorsten Meyer reports encountering hallucinations in personal use. That observation is anecdotal; the source gives no controlled comparative hallucination rates.
Context is capacity, not accuracy
A reported context window of roughly 512,000 tokens describes how much material may fit in a request. It does not show accurate reasoning across all of it.
Cost needs workload data
The report says DeepSeek V4.1 Flash has roughly comparable benchmark intelligence at a much lower measured cost per task, but provides no prices for a numerical comparison.
“I would not choose it for demanding agentic work or long tasks when stronger models are available.”
Thorsten Meyer · ThorstenMeyerAI.comWhat happens next
Weights and workload tests ahead
The benchmark offers a reason for caution. Representative testing can show whether the model fits a team’s actual needs.
API preview available
Announced Oct. 6, 2026. The current evaluation concerns an API-accessible preview.
Public weights planned
Mistral scheduled a weights release later in October, subject to its release terms. The final date and version performance remain unconfirmed in the report.
Test representative tasks
Evaluate the version available to your team and track quality, supervision, and cost.
Measure the work that matters
- Follows constraints across long workflows
- Verifies evidence and claims
- Uses tools correctly across steps
- Completes work with acceptable supervision
- Maintains quality on your own tasks
- Fits your usage and cost patterns
Key questions
Quick answers for developers
Has Mistral Large 4 been released?
It was announced as an API preview on Oct. 6, 2026. Public weights were scheduled for later in October and were not downloadable at the time of the report.
How did it score against competitors?
Artificial Analysis gave the preview 38 in the Oct. 7 comparison. It matched GPT-6 Luna at maximum reasoning effort and ranked below several listed US and Chinese models.
Does the score mean it is unreliable?
No. The index is an aggregate benchmark measure, not a direct reliability test for every task. The report’s hallucination concerns are personal observations, not a controlled study.
What should teams evaluate for agentic work?
Test instruction following, evidence checks, tool use, completion quality, supervision needs, and cost on representative multi-step tasks.
The Benchmark Gap for Mistral
The results matter to developers weighing whether to use Large 4 for complex coding, research or other multi-step work. In the cited index, Claude Opus 5.5 scores 58, Gemini 4 Argon 53 and GPT-6.1 Sol 52. Those are gaps of 20, 15 and 14 index points respectively. They are not percentages and do not predict how often one model will complete a particular task successfully.
For agentic work, a model plans, uses tools and carries decisions across multiple steps. An unsupported assumption early in a workflow can affect later results, even if the final answer sounds coherent. The article’s author argues that this makes dependable execution and verification more important than model size alone. The index offers a broad signal, but does not directly test reliability on every developer’s workload.
Thorsten Meyer reports encountering hallucinations while using the preview and says that experience reduced his confidence in assigning it longer tasks. That is personal experience, not a controlled comparison. It should not be read as evidence that the model always hallucinates more than competitors. Mistral’s advertised strengths in agentic coding and specialized professional tasks also require workload-specific testing.
A large context window does not settle the question. Artificial Analysis reports capacity of roughly 512,000 tokens, meaning that much material can fit in a request; it does not show that a model will reason accurately across all of it. Developers still need to test output quality, supervision needs and cost on the tasks they intend to run.
Popular right now
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
Samsung 42-Inch Class OLED S90F 4K Smart TV (2025)
De'Longhi Magnifica Evo Automatic Espresso Machine with Milk Frother – Silver
As an affiliate, we earn on qualifying purchases.
Preview Now, Weights Later
The current decision concerns an API-accessible preview, not a finished public-weight release. Mistral announced Large 4 on Oct. 6, and its weights were scheduled for release later that month. The company says the model is still being improved, so the preview’s results may not represent the version developers will eventually be able to download.
The benchmark figures are a dated snapshot from Oct. 7, 2026, as reported in the source article. Artificial Analysis’s Intelligence Index aggregates performance across its benchmark suite; the score is not a measure of “intelligence” in every setting, and comparisons using different reasoning-effort settings may not reflect equal compute budgets. Scores may change as models and evaluations are updated.
The source also raises cost as part of the selection decision, reporting that DeepSeek V4.1 Flash delivers roughly comparable benchmark intelligence at a much lower measured cost per task. The supplied material does not include the underlying prices or cost figures, so it does not support a numerical cost comparison here. Developers should compare costs against their own workloads and usage patterns.
“The company says it trained Large 4 on its own infrastructure in Europe and continues to improve it.”
— Mistral
What the Preview Cannot Establish
The available comparison does not show how Large 4 performs across every coding, research or professional workflow, or whether its advertised strengths hold up under independent, task-specific testing. A score of 38 does not prove the model will fail a particular assignment. The source also does not provide controlled comparative hallucination rates; Meyer’s observations are anecdotal.
The model’s publicly downloadable weights had not yet been released at the time covered by the report, and the eventual release date and final version’s performance remain to be confirmed. The supplied material also gives no detailed price table, so the cost comparison with DeepSeek cannot be independently quantified from these figures. Results may shift as Mistral updates the model and benchmark scores change.
Weights and Workload Tests Ahead
The next stated milestone is Mistral’s planned release of Large 4’s weights later in October. Developers can then assess the publicly available model alongside the API preview, subject to Mistral’s release terms. The company says it is continuing to improve the model, but the source does not give a more specific update schedule.
For teams considering adoption, the practical next step is to test the version they can access on representative tasks: measure whether it follows instructions across long workflows, checks evidence, uses tools correctly and completes work with an acceptable level of supervision. They should also track cost for those tasks. Until such results are available, the benchmark offers a reason for caution, not a definitive ranking for every use case.
Key Questions
Has Mistral Large 4 been released?
Mistral announced it as an API preview on Oct. 6, 2026. The model weights were scheduled for release later in October and were not publicly downloadable at the time of the report.
How did Mistral Large 4 score against competitors?
Artificial Analysis gave the preview an Intelligence Index score of 38 in the Oct. 7 comparison. That was below several listed US and Chinese models, while matching GPT-6 Luna at maximum reasoning effort.
Does the score mean Mistral Large 4 is unreliable?
No. The index is an aggregate benchmark measure, not a direct test of reliability on every task. The source’s author reports personal concerns about hallucinations, but does not cite a controlled comparative study.
Is Mistral Large 4 open weight now?
Not according to the report. At the time covered, developers could access the preview through an API; Mistral had scheduled the weights for a later release in October.
What should developers evaluate before using it for agentic work?
Teams should test the available version on their own multi-step tasks, including how well it follows constraints, verifies evidence and uses tools, as well as how much supervision and cost it requires.
Source: ThorstenMeyerAI.com
