Alibaba’s AI tops global math contests, challenging OpenAI

Alibaba has unveiled Qwen3-Max-Thinking, an upgraded AI model that achieved perfect scores in the 2025 American Invitational Mathematics Examination (AIME) and the Harvard-MIT Mathematics Tournament (HMMT).

Alibaba said this is the first time a Chinese AI reasoning model has reached 100% accuracy in these contests.

These competitions are considered among the most difficult benchmarks for AI problem-solving and reasoning abilities.

OpenAI’s GPT-5 Pro has also self-reported perfect scores in both events.

Qwen3-Max-Thinking was built using Qwen3-Max, Alibaba’s largest AI model, which has over 1 trillion parameters and was unveiled in late September.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Benchmark wins lack verification

Perfect scores on American Invitational Mathematics Examination (AIME) and Harvard-MIT Mathematics Tournament (HMMT) come from the vendor with no third-party check or public leaderboard.
Missing details cover whether the test used 2025 problems, a closed-book run (no internet or tool access during testing), and contamination controls (checks to ensure test questions were not in training data).
A two-week cryptocurrency trading test used Qwen3-Max’s 22.3% return versus GPT-5 Pro’s 62.7% loss, a window too short for statistical weight so luck can swamp skill.
Independent, reproducible evidence with clear methods is still missing, so breakthrough claims for Qwen3-Max-Thinking on reasoning look early.

API access opens routing options, pricing and regions matter

Software and data teams can try cost-performance routing by sending requests to Qwen3-Max-Thinking or other providers by price or accuracy to trim inference costs (the cost of generating model outputs) 1.
Asia-Pacific (APAC) builders get a local option but still need API pricing plus regional availability beyond Singapore and quota limits with latency benchmarks for unit economics (per-request costs and margins) 1 2.
Investors in AI infrastructure should watch adoption signals. If developers route complex reasoning through Qwen’s API while using Western providers for other work, that hints at competitiveness beyond benchmarks.
Builders need batch inference availability (processing many requests to lower cost), context caching discounts (reusing previously processed tokens to avoid recomputation), and clarity on whether pay-per-token model (billing per unit of text) for multi-turn conversations (back-and-forth exchanges across messages) stays cost-effective versus rivals at scale 2 3.