GPT5 was released last night, the impact on $NVIDIA(NVDA.US) is still uncertain, but I've seen some reviews, sharing here:

#Points that exceeded expectations

1) Significantly reduced hallucinations, which is currently the biggest issue I face when using AI tools

During online searches, GPT-5's answers had 45% fewer factual errors compared to GPT-4o. When thinking independently, the error rate was 80% lower than GPT-3

2) Pricing is lower than expected

GPT5 is available to all users including free users, Pro users get access to GPT-5 Pro (smarter version)

A price interface input $1.25/million tokens, output $10/million tokens

#Points that met expectations (especially compared to leaks from information sources)

1) Code capabilities have indeed improved

SOTA in SWE-Bench, SWE-Lancer, Aider Polyglot. Human final test 42%, SWE 75%.

Based on subsequent tests, coding ability has indeed improved, #some tasks surpassed Claude

2) Math ability improved to AIME 94.6%, reasoning ability improved to GPQA 88.4% reaching SOTA

3) Unified model entry point, GPT5 now decides when to activate deep thinking, previous versions were considered too complex

#Points that fell short of expectations

1) Didn't surpass grok4 on ARC-AGT-2 LEADERBOARD, only slightly better than Claude Opus 4

(ARC tasks cover various abstract logics and thinking modes, including IQ tests that humans solve easily but LLMs previously performed poorly on)

2) Multimodal capabilities only significantly improved voice, some expected coherent video input capability

3) Knowledge cutoff is 2024, not updated to latest 2025 information

4) According to some subsequent evaluations, creative writing ability is worse than previous models, instruction following is mediocre