\n\n
Key developments this month show a clear shift in the frontier of reasoning AI. Anthropic’s Claude Sonnet 5 achieved a perfect 36‑out‑of‑36 on an IQ‑style test, scoring 145 and outpacing the next best model by 24 points, while OpenAI’s ChatGPT recorded 121 【Claude Sonnet 5 tops the field at 145, posting a perfect 36/36 and beating every other tested model】【ChatGPT trails the field at 121 — putting a 24-point gap between the top and bottom scorers】. This marks the widest performance gap recorded in the latest benchmark cycle.
Anthropic also rolled out Claude Fable 5.1 on September 1, 2026, making it instantly available across all platforms, and Google responded with Gemini 3.8 Flash on September 2, 2026. Both models target the high‑effort reasoning market, but early results show differing strengths: Claude Fable 5.1 leads the CursorBench 4.0 agentic‑coding benchmark with a 51.8 % success rate, though at a high cost of $17.28 per task, whereas xAI’s Grok 4.7, while still behind on general reasoning, remains competitive on narrower agentic tasks and costs far less per task【Anthropic released Claude Fable 5.1 on September 1, 2026, generally available on every platform on day one with no staged preview】【Google followed a day later with Gemini 3.8 Flash reaching general availability on September 2】【Claude Fable 5.1 (max effort) leading at 51.8% on CursorBench 4.0 but at a steep cost of $17.28 per completed task】【Grok 4.7 is still catching up on general reasoning even where it’s competitive on narrower agentic tasks】.
Reliability remains a differentiator. On the FACTS grounding benchmark, GPT‑6 Astra posted the highest score at 73.1 %, while Claude Opus 4.7 kept hallucinations low at 36 % compared with Grok 4’s 64 %【On FACTS Grounding, GPT-6 Astra leads at 73.1% and Claude Opus 4.7 hallucinating on just 36% of questions versus 64% for Grok 4】. Meanwhile, GPT‑6 Astra also demonstrated superior cost efficiency, outscoring Claude Opus 5 on AutomationBench while costing only 9 % of its price per completed task【GPT-6 Astra outscored Claude Opus 5 on AutomationBench while costing just 9% of Opus 5’s price per completed task】.
Why it matters: The widening performance gap gives enterprises clearer guidance on which model to choose for high‑stakes reasoning tasks, while cost‑per‑task and hallucination data help balance budget constraints against reliability needs. The rapid succession of releases—Claude Fable 5.1, Gemini 3.8 Flash, and Grok 4.7—means the market will continue to refine the trade‑offs between raw reasoning power, economic efficiency, and factual accuracy.
What to watch next:
1. Upcoming updates to Claude Fable 5.1 and Gemini 3.8 Flash that may improve cost‑efficiency without sacrificing accuracy.
2. The next round of FACTS and Intelligence Index benchmarks, where OpenAI’s GPT‑6 line aims to close the hallucination gap further.
3. Real‑world deployments that stress‑test these models in high‑risk domains (finance, healthcare, autonomous systems), which will reveal whether the current benchmark advantages translate into operational safety. [[1]]