AI ranking 2026 : les meilleurs modèles et agents

The ranking draws on several public benchmarks to compare models by use: conversation, code, agents, cost and performance. The Le Recul score is the final composite score calculated from those sources — not a source added to the calculation.

The Le Recul score is indicative and composite: it aggregates several public sources (LM Arena, SWE-bench, Aider Polyglot, and Artificial Analysis once the API is connected), normalises the values on a 0–100 scale and applies a documented weighting per category. This score is not a source added to the calculation — it is the result of the aggregation. No data is invented: if a category lacks measurements, the display says “insufficient data”.

The Assessment column reflects source coverage for each model: Reliable when at least two relevant independent sources confirm the model (cross-validation), or when a strong reference source (LM Arena, Artificial Analysis, LiveBench, Epoch) on its own covers at least 60 % of the category's weighted criteria — preview, beta and experimental models, along with effort or reasoning variants, are excluded from that promotion and stay “Partial”; Partial when a single source documents the model with usable but limited coverage — often a recent or single-source model, not yet cross-checked by a second independent source; Insufficient when the data does not allow a clean verdict. A knowledge measurement alone (MMLU-Pro) is never enough to make a model “Reliable”, and categories where no strong source covers enough criteria (for example Code) stay “Partial” until a second independent source is available.

Le Recul score. The score shown is the settled Le Recul score: it combines the scores of reliable public sources (LM Arena, SWE-bench, Aider, MMLU-Pro, BBEH, LiveBench, Artificial Analysis once the API is connected) according to a documented weighting per category. Internally, a tiny tie-breaking component (≤ 0.01) folds in coverage, the number of independent sources, the assessment status and how recent the data is — invisible at two decimal places for normal scores, it separates models that look equivalent. The score shown is therefore a Le Recul score, not a raw average: apparent ties are settled by coverage, stability and relevance of the sources. The score shown is limited to two decimal places. No model is removed or hidden in the event of a tie.

The Cost / performance category compares overall performance with the price of the public monthly subscription (Claude.ai Pro, ChatGPT Plus, Gemini Advanced, Le Chat Pro, free interfaces…). API prices and per-token costs are not taken into account, because they do not reflect what an ordinary user actually pays.

The Local / open-source AI category gathers open-weights or open-source models that can be used locally without heavy infrastructure. Very large models (beyond roughly 35 billion parameters) are excluded from the main local ranking by editorial choice. Proprietary cloud models (Claude, GPT, Gemini, Grok, Mistral cloud…) never appear in this category.

The Coding agents category gathers agentic systems (scaffolds + models) measured on complete coding tasks: SWE-bench (solving real bugs) and Aider in architect mode. Models on their own (Aider whole/diff) stay in the Code category.

Additional components. BBEH (DeepMind's official leaderboard) feeds the reasoning criterion; MMLU-Pro (TIGER-Lab's official dataset) feeds an internal knowledge criterion (14 disciplines). Both benchmarks count as independent sources for the Le Recul score — never displayed as a separate public column, and never counted twice. Their effect stays measured: they mainly help to cross-check models already covered by other sources; a recent model that MMLU-Pro does not yet cover stays “Partial”. Benchmarks for which no usable source has been identified (IFEval, classic BBH, MATH, GPQA, MuSR) are switched on as soon as the Artificial Analysis connection is available.

Hidden categories. Some planned categories (agent families in particular) may be temporarily hidden from the public ranking as long as no source provides enough rankable models. They are not deleted: they reappear automatically as soon as a collection run provides sufficient data. A category that contains models is never hidden, even if many of its assessments are still “Partial”.

Data continuity. If a source does not respond during a cycle, the ranking is not emptied: that source's last known reliable measurement is kept, within a freshness limit. A guardrail blocks any publication that would abruptly degrade the ranking (a critical source lost, an abnormal drop). No data that is too old is presented as fresh.

An indicative score, weighted from public sources. Transparent methodology.

Cadence : Mise à jour 1×/jour, heure variable. The site is not rebuilt on every cycle: only the data files are updated.

See the full FAQ →