In the Artificial Analysis ranking consulted on October 28, 2026, the best model whose weights are downloadable obtains 32 points, compared to 58 for Claude Opus 5.5: the difference was 12 points in the previous version of the index. Claude Haiku 5.5, Anthropic’s closed model, gets 36. In this 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 assessments with sets of questions that laboratories cannot consult.
ActuIA has brought the Artificial Analysis open models page, still displayed in version 4.3.2 on October 28, closer to the new ranking. The eleven models that this page lists all lose between 12 and 16 points. Mistral Large 4, whose weights have not yet been published, loses 13. Claude Opus 5.5 obtained 58 points on October 23 and still obtains 58 in version 5. The figure that Artificial Analysis puts forward, a gap of 12 points between open and closed models, counts Kimi K3.1 Pro among the open models even though its weights are not published.
60% of the index is now based on private tests
According to the version 5 overview article, 60% of the index is based on assessments whose test sets are private or whose answers are kept secret. The index is divided into four families, each weighing 25%: intellectual work, code, sciences, general knowledge. Artificial Analysis cites as private FrontierCode 2.0, a programming benchmark launched by Cognition whose tasks “remain private to avoid contamination”, and a subset of 40 difficult questions from CritPt v2, in physics. AA-Omniscience v1.1 has been purged of its public issues. The article does not detail the exact composition of the 60%.
Three tests come out of the index: GDPval-AA, SciCode and AA-LCR. Two are entering: FrontierCode 2.0 and Terminal-Bench-Science, scientific tasks carried out in a terminal. HLE-Diamond replaces Humanity’s Last Exam. On September 4, on its X account, Artificial Analysis already increased the private share to 40% with version 4.2, to prevent the optimization of models for classification, and announced an increase in version 5.
Twelve open models lose between 12 and 16 points
The table compares the scores of the open models page (version 4.3.2) and those of the ranking in version 5, recorded on October 28. For Mistral Large 4, absent from the open page, the score 4.3.2 comes from its file and the ActuIA article of October 19.
| Model | Index 4.3.2 | Index 5 | Écart |
|---|---|---|---|
| MiMo-V2.6-Pro (Xiaomi) | 46 | 30 | −16 |
| DeepSeek V4.1 Flash 1017 (max) | 45 | 32 | −13 |
| GLM-5.3 (max) | 45 | 32 | −13 |
| Ling 3.1 Flash 1019 (Ant Group) | 44 | 29 | −15 |
| Step 5 Preview | 44 | 29 | −15 |
| Muse Spark Open (Meta, xhigh) | 39 | 25 | −14 |
| Mistral Large 4 Preview | 38 | 25 | −13 |
| Qwen3.8 27B (xhigh) | 34 | 20 | −14 |
| K2 Horizon 375B A23B | 31 | 18 | −13 |
| Beam 1.0 (Reflection AI) | 30 | 18 | −12 |
| Nemotron 3.5 Super VL (NVIDIA) | 26 | 14 | −12 |
| Inkling (xhigh) | 25 | 12 | −13 |
Kimi K3.1 Pro from Moonshot AI, whose weights are not published, goes 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. He only mentions contamination, i.e. the presence of test questions in the training data, in relation to FrontierCode 2.0. The new composition also weighs. Sciences are worth a quarter of the index, and on Terminal-Bench-Science, GPT-6 Astra succeeds 63% of the tasks compared to 13% for DeepSeek V4.1 Flash. Without trial-by-trial scores for both versions, the share of each effect cannot be calculated.
3, 12 or 26 points depending on the model counted as open
Artificial Analysis summarizes version 5 with one sentence: the gap between open weight models “goes from 3 to 12 points” (translated from English). This calculation takes Kimi K3.1 Pro as the best open model, at 46 points, specifying that its weights are “not yet published”. The starting 3 was already based on this choice. In its October 23 article on Kimi K3.1 Pro, Artificial Analysis wrote that the gap would be reduced “from 12 points to 3” once its weights were published. They still were not there on October 28.
The Kimi K3.1 Pro sheet on the same site also classifies it as a “Proprietary model”. Mistral Large 4, in the same situation, is also classified as owner and absent from the open models page. On the weights actually published, Artificial Analysis gives the difference lower 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 ahead of it by one point on October 19. Both are the highest ranked non-Chinese models among those with published or announced weights. DeepSeek V4.1 Flash and GLM-5.3 are ahead of 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 there.
The Large 4 sheet, upgraded to version 5 on October 28, does not display its results event by event. We therefore do not know if it maintains the lead it had in version 4.3.2 on the 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 low effort model does better than all published weights
The ranking 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. This is 3 points more than the best model at published weights. At maximum effort, the same task costs him $10.04. A purchase comparison between a downloadable model and an API model therefore also concerns the force setting chosen.
The Hugging Face depot of Mistral Large 4 displays as the estimated date of publication of the weights October 31, 2026, without file or license as of October 28 in the evening. Mistral wrote on October 6 (see today’s ActuIA analysis) that he would publish them “by the end of the month.” The date of October 27, put forward by The Next Web and The New Stack without a source, was not kept. Two pieces of information are missing to place Large 4 among the open models: the license chosen, and the score of the published version, whose training by reinforcement continued according to the Mistral post of October 6.
Updated October 30, 2026. Mistral released the Large 4 weights on October 29 under a modified MIT license, which imposes a commercial license on companies operating a model offering as a service. The final version scores 26 points on version 5 of the index, one more than the preview, and is now listed on Artificial Analysis’s open weight models page at number seven, ahead of Muse Spark Open (25). The details appear in the updated ActuIA analysis.
Our articles will be prioritized in Google’s “Featured” section.
