@maiff20: Gave the 2025 college entrance exam math paper to Codex, told it not to access the internet, the reasoning process was actually using Python to compute, rather than looking up historical information in the model, finished in 7 minutes, scored 150 points, and even pointed out contradictions in the answers.
Summary
A user tested OpenAI Codex on the 2025 college entrance exam math paper, restricting it from accessing the internet and only using Python for real computation. It completed the test in 7 minutes, scored a perfect 150 points, and also pointed out contradictions in the official answers.
View Cached Full Text
Cached at: 06/08/26, 11:24 AM
I gave the 2025 Gaokao math exam to Codex, asked it not to access the internet, and the thinking process was actually done using Python rather than searching for historical information in the model. It finished in 7 minutes, scored 150 points, and even pointed out the contradictory parts in the answer key. https://t.co/qjZThHx5KH
Similar Articles
Meet a mathematician solving previously unsolvable math problems with GPT-5.6
A mathematician used Codex 5.6 to successfully disprove an algebraic surface conjecture that had taken him three years to try to prove. The model can automatically spawn sub-agents to handle heavy computations, allowing him to focus on difficult problems and life.
@AISuperDomain: OpenAI is specifically screwing over domestic users! It turns out that many Codex users who set the reasoning level to Ultra are actually assigned a downgraded gpt-5.5-mini model. But most users haven't noticed at all — they just mindlessly run loops. However, the detection method is also simple — you can use the prompt below to test whether it is...
The post claims that OpenAI Codex assigns a downgraded gpt-5.5-mini model to domestic users at the Ultra reasoning level, provides a detection prompt, and calls on users to reply 1 or 2 to measure the downgrade rate for domestic users.
@MaxForAI: A brutal fact: AI's mathematical abilities have already surpassed 99% of humans on this planet. @MenloVentures partner Deedy @deedydas, who invested in Anthropic and OpenRouter, conducted a test. He used the just-concluded 2026 International Mathematical Olympiad...
AI models Fable, Sol, K3, and Axiom all achieved a perfect score of 42/42 in the 2026 International Mathematical Olympiad, solving the competition completely for the first time at low cost. Among them, Claude Fable 5 was the fastest, while GPT 5.6 Sol had the lowest cost.
@FakeMaidenMaker: This is probably the most complete Codex practical course yet — author Riley Brown spends 30 minutes walking through 7 real white-collar work scenarios, showing you one by one how to boost efficiency with Codex: - Auto OCR 60 receipt photos into Excel - Scan two weeks of emails and organize…
This content introduces the seven major capabilities of the AI agent Codex, demonstrating how to improve work efficiency through 7 real white-collar work scenarios (such as OCR receipts, organizing emails, creating PPTs, etc.), making it a complete practical tutorial.
@amiaoapp: 70 years of Gaokao math exam questions have been neatly compiled in a single GitHub repository. From 1952 to 2025, nearly full coverage. To put it dramatically, it's like packaging the historical data of Gaokao math into a database. The project is called gaokaomath, and its features are straightforward: all past Gaokao math exam questions …
The gaokaomath project on GitHub has compiled all Gaokao math exam PDFs from 1952 to 2025, clearly categorized by year, continuously updated, convenient for practice and teaching research.