Leaked Gemini 4 Neon benchmarks show AA Index 67, topping Claude Fable 5.2 across the board
kimmonismus · x · 2026-11-18
A benchmark table circulating for the unannounced "Gemini 4 Neon" (sourced to a deepmind.google evals-methodology page, unverified) shows it leading every listed model: AA Index v5.0.1 67 vs Fable 5.2's 65, SWE-sweep 15.8%, Terminal-Bench 4.1 (cheating-adj.) 71.4%, Agents' Last Exam 81.6%, AutomationBench-AA 82.4%, Vending-Bench 2 bank balance $18,920, Chartography 76.8%, HLE-Diamond 83.4%. Its only loss: OSWorld-2.1 offline at 84.6% vs Claude Opus 5.5's 87.2%. None of the model names (Gemini 4 Carbon/Hydrogen, GPT-6.2 Astra, Claude Fable 5.2) are officially confirmed; treat as rumor until an official release.
More from Models
- Epoch: GPT-6.2 does 12.3 serial reasoning steps per token, raising monitorability concerns — austinc3301 · 2026-11-18
- $20 Claude Opus 5.5 vs free Hy4 Flash: one-shot dance test showdown — CodeByPoonam · 2026-11-18
- History repeats: the loudest critics pan new models, real users stay quiet — daniel_mac8 · 2026-11-18
- SmolVLM2-256M-Video-Instruct trends on Hugging Face — HuggingFaceTB · 2026-11-18
- FluidVoice fluid-speech-pro ASR model trends on Hugging Face — FluidVoice-Labs · 2026-11-18
- Perplexity quietly ships pplx-decider-v1.3, a Qwen3.8-27B fine-tuned classifier on Hugging Face — perplexity-ai · 2026-11-18