Mistral Large 4 drops from 38 to 25 on new Artificial Analysis index

On the Artificial Analysis leaderboard consulted on 28 October 2026, the best model with downloadable weights scores 32 points, against 58 for Claude Opus 5.5: the gap was 12 points in the previous version of the index. Claude Haiku 5.5, Anthropic's closed model, scores 36. In that previous version, still online a few days earlier, Xiaomi's MiMo-V2.6-Pro, DeepSeek V4.1 Flash and GLM-5.3 all exceeded 44 points. The models have not changed. The new version of the index replaces public evaluations with question sets that labs cannot access.

ActuIA compared the open models page of Artificial Analysis, still displayed in version 4.3.2 on 28 October, with the new leaderboard. The eleven models listed on that page all lose between 12 and 16 points. Mistral Large 4, whose weights are not yet published, loses 13. Claude Opus 5.5 scored 58 points on 23 October and still scores 58 in version 5. The figure Artificial Analysis highlights, a 12-point gap between open and closed models, counts Kimi K3.1 Pro among open models even though its weights are not published.

60% of the index now relies on private tests

According to the presentation article for version 5, 60% of the index relies on evaluations whose test sets are private or whose answers are kept secret. The index is divided into four families each weighing 25%: intellectual work, code, science, general knowledge. Artificial Analysis cites as private FrontierCode 2.0, a programming benchmark launched by Cognition whose tasks "remain private to avoid contamination" (translated from English), and a subset of 40 difficult questions from CritPt v2, in physics. AA-Omniscience v1.1 has been purged of its public questions. The article does not detail the exact composition of the 60%.

Three tests leave the index: GDPval-AA, SciCode and AA-LCR. Two enter: FrontierCode 2.0 and Terminal-Bench-Science, scientific tasks carried out in a terminal. HLE-Diamond replaces Humanity's Last Exam. On 4 September, on its X account, Artificial Analysis had already raised the private share to 40% with version 4.2, to prevent models from being optimised for the leaderboard, and announced an increase in version 5.

Twelve open models lose between 12 and 16 points

The table compares the scores from the open models page (version 4.3.2) and those from the version 5 leaderboard, recorded on 28 October. For Mistral Large 4, absent from the open page, the 4.3.2 score comes from its model card and from ActuIA's article of 19 October.

Moonshot AI's Kimi K3.1 Pro, whose weights are not published, drops from 55 to 46: 9 points less. The best-ranked closed models remain between 50 and 58 points in version 5: Claude Opus 5.5 (58), GPT-6 Astra (57), GPT-6.1 Sol (55), Claude Sonnet 5.5 (54), GPT-6.1 Terra (51) and Gemini 4 Argon (50).

Artificial Analysis does not explain the fall of open models. It mentions contamination, i.e. the presence of test questions in training data, only regarding FrontierCode 2.0. The new composition also weighs. Science accounts for a quarter of the index, and on Terminal-Bench-Science, GPT-6 Astra succeeds in 63% of tasks against 13% for DeepSeek V4.1 Flash. Without the test-by-test scores of both versions, the share of each effect cannot be calculated.

3, 12 or 26 points depending on the model counted as open

Artificial Analysis summarises version 5 with one sentence: the gap of open-weight models "goes from 3 to 12 points" (translated from English). This calculation takes as best open model Kimi K3.1 Pro, at 46 points, noting that its weights are "not yet published". The initial 3 was already based on this choice. In its article of 23 October on Kimi K3.1 Pro, Artificial Analysis wrote that the gap would narrow "from 12 points to 3" once its weights were published. They still were not on 28 October.

The Kimi K3.1 Pro model card on the same site actually classifies it as "Proprietary model". Mistral Large 4, in the same situation, is also classified as proprietary and absent from the open models page. On the only weights actually published, Artificial Analysis gives the gap lower down in its article: DeepSeek V4.1 Flash and GLM-5.3, at 32, are 26 points behind Opus 5.5. Before the version change, it was 12 points. The gap has therefore more than doubled.

Mistral Large 4 tied with Muse, 7 points behind DeepSeek and GLM

With 25 points, Mistral Large 4 joins Meta's Muse Spark Open, which was one point ahead on 19 October. The two are the best-ranked non-Chinese models among those whose weights are published or announced. DeepSeek V4.1 Flash and GLM-5.3 lead them by 7 points, MiMo-V2.6-Pro by 5, Ling 3.1 Flash and Step 5 by 4. No other European model has a score in version 5: Mistral Medium 3.5 and Mistral Small 4 do not appear.

The Large 4 model card, switched to version 5 on 28 October, does not display its test-by-test results. It is therefore unknown whether it retains the lead it had in version 4.3.2 on agent tests: 27% success against 8% for Muse on Terminal-Bench 4.0, 60% against 50% on AutomationBench. These two tests remain in version 5, with 15% of the total weight.

A closed model at low effort beats all published weights

The leaderboard indicates the reasoning effort setting of each measured model: max, xhigh or high. Set to low effort, Claude Opus 5.5 scores 35 points for an average cost of $0.98 per index task, according to Artificial Analysis. That is 3 points more than the best model with published weights. At maximum effort, the same task costs it $10.04. A purchase comparison between a downloadable model and an API model therefore also concerns the effort setting chosen.

The Hugging Face repository of Mistral Large 4 shows an estimated weight release date of 31 October 2026, with no file or licence as of the evening of 28 October. Mistral had written on 6 October (see ActuIA's analysis of that day) that it would publish them "by the end of the month". The date of 27 October, put forward by The Next Web and The New Stack without source, was not met. Two pieces of information are missing to place Large 4 among open models: the chosen licence, and the score of the published version, whose reinforcement learning training was continuing according to Mistral's post of 6 October.

Add ActuIA to my preferred sources on Google

Our articles will then appear first in Google Top Stories.