AI Agents: Claude Fable 5.1 Outperforms GPT-6 Astra in September

Key Takeaways

1Claude Fable 5.1 takes the lead in the agentic ranking of Arena ahead of GPT-6 Astra

2The ranking is based on five signals derived from nearly 1.7 million sessions across 43 models

3Anthropic holds six out of ten positions, OpenAI three, while the rest of the pack sees Tencent, DeepSeek, Z.ai, and Meta making progress

Why it matters : This ranking, based on real sessions, highlights the diversity of model strengths and the rapid evolution of the agentic landscape.

The agentic ranking from Arena in mid-September 2026 places Claude Fable 5.1 at the top, ahead of GPT-6 Astra. The verdict aggregates five signals derived from nearly 1.7 million sessions across 43 models, observed in real-world conditions in Agent Mode. Anthropic occupies six out of ten spots, but OpenAI is climbing and takes second place.

Different leaders depending on the measured signals

No model dominates all five aggregated criteria. Claude Fable 5.1 achieves the highest rate of tasks validated by users. It is GPT-6 Astra that shows the largest share of positive feedback. The two editions of Claude Opus 5 stand out for their integration of corrections and ability to get back on track after a failed command. Among the top ten, only GPT-6 Astra presents a negative steerability score, an indicator of the model's reaction when a user corrects it. The five signals considered cover user validation or rejection, natural language ratings, integration of corrections, getting back on track after a failed command, and invoking unavailable tools.

How Arena evaluates models in agent mode

Arena's agentic ranking does not include anonymized duels or the Elo score from its overall ranking. The platform uses a methodology known as causal tracing. Performance is observed in Agent Mode, a sandbox opened in June where the user delegates a complete workflow to an agent capable of mobilizing multiple tools. The model in control changes randomly between sessions to isolate its contribution to the outcome. The current verdict is based on nearly 1.7 million sessions covering 43 models.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

A top 10 dominated by Anthropic, with OpenAI lurking

Claude Fable 5.1, available in Max version since September 1, establishes itself immediately upon launch. GPT-6 Astra, which was made public two days later and is the first OpenAI model to be listed at the highest cyber risk level, ranks in second place. The two variants of Claude Opus 5 occupy the third and fourth positions, followed by Claude Fable 5 and Claude Opus 4.8. Anthropic holds six of the top ten spots, while OpenAI has three, including Claude Sonnet 5 in ninth place and GPT-5.6 Sol in seventh after a third place finish in August. Kimi K3 drops from fifth to eighth place. The complete order of the top ten is published, with GPT 5.5 in tenth position. The previous month, OpenAI's best model was only fourth.

Beyond the podium: a tight pack and setbacks

Hy4 preview from Tencent enters the ranking in eleventh place, ahead of DeepSeek V4.1 Flash, GLM 5.2 (Z.ai), and Muse Spark 1.3, which represents the latest version of Meta's agentic model. In August, it took reaching the thirteenth place to identify a player different from Anthropic, OpenAI, or Moonshot AI, and the twenty-second to see Meta appear. Google remains distant with Gemini 3.6 Flash in thirty-fourth position. Mistral ranks forty-first, the only European player in this table.