33-benchmark showdown: Claude Fable 5.2 leads overall, GPT-6.2 Astra close behind
HealthySkeptic2000 · reddit · 2026-11-21
A Reddit user compiled a 33-benchmark comparison of four flagship models as of Nov 21, 2026.
Key points:
- Overall indexes: Claude Fable 5.2 leads AA Intelligence Index (64.6), Vals Index (71.31%), and ties LiveBench 2.0 (84.9) with GPT-6.2 Astra; Astra tops Epoch Capabilities Index (174.2).
- Coding: Fable sweeps most—SWE Atlas 80.5%, Vibe Code Bench 93.8%, WebDev Arena Elo 1,824; Kimi K3.1 Pro only wins Terminal-Bench 4.0 AA (67.2%).
- Agents: split results—Astra takes Agents' Last Exam (78.3%), Gemini 4 Carbon takes APEX-Agents (84.2%), Fable leads Coarena Computer Use (Elo 1,104), Kimi tops GDP.pdf (31.4%).
- Reasoning: Fable dominates HLE-Diamond (80.2%) and SimpleQA (94.0%); Kimi trails badly overall (AA index 46.5).
- Finance & law: Gemini 4 Carbon leads Vals Legal Index (75.93%) and Tax Agent (51.8%).
Data drawn from Artificial Analysis, Vals AI, Epoch AI and other third-party leaderboards with independent methodologies.
More from Models
- Inkling 2 mini impresses as a local model on M5 Max, token-efficient at low effort — LiTianleli · 2026-11-21
- Thinking Machines ships Inkling 2 Mini: 26B open-weights model that runs on a laptop — ziqiao_ma · 2026-11-21
- Users push Google to open Gemini subscriptions to third-party agent harnesses — WolframRvnwlf · 2026-11-21
- Reddit user buys $200 Claude Max for $29, wonders how resellers profit — mehdiweb · 2026-11-21
- OpenRouter exposes BYOK usage; top model Hy4 accused of faking volume — gajesh · 2026-11-21
- All six remaining Millennium Prize Problems see claimed breakthroughs since August 2026, led by OpenAI and Anthropic — haider1 · 2026-11-21