Wednesday, 17th September 2025
In July it was the International Math Olympiad (OpenAI, Gemini), today it's the International Collegiate Programming Contest (ICPC). Once again, both OpenAI and Gemini competed with models that achieved Gold medal performance.
OpenAI's Mostafa Rohaninejad:
We received the problems in the exact same PDF form, and the reasoning system selected which answers to submit with no bespoke test-time harness whatsoever. For 11 of the 12 problems, the system’s first answer was correct. For the hardest problem, it succeeded on the 9th submission. Notably, the best human team achieved 11/12.
We competed with an ensemble of general-purpose reasoning models; we did not train any model specifically for the ICPC. We had both GPT-5 and an experimental reasoning model generating solutions, and the experimental reasoning model selecting which solutions to submit. GPT-5 answered 11 correctly, and the last (and most difficult problem) was solved by the experimental reasoning model.
And here's the blog post by Google DeepMind's Hanzhao (Maggie) Lin and Heng-Tze Cheng:
An advanced version of Gemini 2.5 Deep Think competed live in a remote online environment following ICPC rules, under the guidance of the competition organizers. It started 10 minutes after the human contestants and correctly solved 10 out of 12 problems, achieving gold-medal level performance under the same five-hour time constraint. See our solutions here.
I'm still trying to confirm if the models had access to tools in order to execute the code they were writing. The IMO results in July were both achieved without tools.
Anthropic: A postmortem of three recent issues. Anthropic had a very bad month in terms of model reliability:
Between August and early September, three infrastructure bugs intermittently degraded Claude's response quality. We've now resolved these issues and want to explain what happened. [...]
To state it plainly: We never reduce model quality due to demand, time of day, or server load. The problems our users reported were due to infrastructure bugs alone. [...]
We don't typically share this level of technical detail about our infrastructure, but the scope and complexity of these issues justified a more comprehensive explanation.
I'm really glad Anthropic are publishing this in so much detail. Their reputation for serving their models reliably has taken a notable hit.
I hadn't appreciated the additional complexity caused by their mixture of different serving platforms:
We deploy Claude across multiple hardware platforms, namely AWS Trainium, NVIDIA GPUs, and Google TPUs. [...] Each hardware platform has different characteristics and requires specific optimizations.
It sounds like the problems came down to three separate bugs which unfortunately came along very close to each other.
Anthropic also note that their privacy practices made investigating the issues particularly difficult:
The evaluations we ran simply didn't capture the degradation users were reporting, in part because Claude often recovers well from isolated mistakes. Our own privacy practices also created challenges in investigating reports. Our internal privacy and security controls limit how and when engineers can access user interactions with Claude, in particular when those interactions are not reported to us as feedback. This protects user privacy but prevents engineers from examining the problematic interactions needed to identify or reproduce bugs.
The code examples they provide to illustrate a TPU-specific bug show that they use Python and JAX as part of their serving layer.