An unannounced Flash model that users are calling Gemini 3.2 Flash is drawing attention for one narrow but important reason: community testers say it can solve IMO 2025 Problem 6, one of the hardest public math problems now being used to probe frontier AI reasoning.
That matters because Problem 6 is the kind of question that usually separates competent mathematical fluency from genuinely unusual reasoning performance. The current wave of benchmarking around the International Mathematical Olympiad suggests Google may be pushing its Flash line harder than most companies push their fast models, and the result is forcing a fresh look at what "reasoning" means when it shows up in a speed-tier AI system.
The claim is circulating through active Reddit testing, where users say Gemini 3.2 Flash can solve IMO 2025 Problem 6 and that GPT-5.5 Pro is the only other publicly tested model they have seen match it without extra scaffolding. That distinction matters. This is not an official Google benchmark or a lab paper. It is a live community claim around an apparently unreleased or quietly routed model, and it should be treated with the right amount of caution.
Still, the timing makes the story relevant. Fresh reports over the past few days have pointed to Gemini 3.2 Flash appearing in app builds, AI Studio metadata, or hidden routing behavior ahead of Google I/O 2026. Google has not formally announced the model, but the discussion is current, and the IMO problem itself is legitimate. Google DeepMind has already published its own IMO 2025 solution PDF, confirming the problem set and showing why these tasks have become a useful stress test for mathematical reasoning.
IMO Problem 6 is not just another hard math prompt. In olympiad circles, the final problem is designed to resist ordinary techniques, and even strong solvers can spend hours without finding a clean path through it. That is why AI performance on a problem like this gets treated as a signal, not just a score. If a model can produce a valid route through the last problem of the IMO set, it suggests something more structured is happening, at least in that narrow setting.
Google's own IMO materials show how much discipline is required even for a successful model-written proof. The company's published solutions describe rigorous multi-step reasoning across geometry, algebra, and combinatorics, the sort of output that makes these benchmarks useful but also easy to overread. A model that can solve one exceptional problem is not suddenly a mathematician, but it is no longer fair to dismiss the result as a parlor trick either.
Flash models are climbing
The more interesting angle is that this would be a Flash-tier model, not a slower premium reasoning system. That fits the direction Google has already taken with the Gemini lineup. As Google's December developer post described Gemini 3 Flash, it was built as "frontier intelligence built for speed," a phrase that made clear the company wanted Flash to be more than a lightweight fallback.
If the Gemini 3.2 Flash reports hold up, Google is compressing capability into a cheaper and faster tier faster than many rivals expected. That is commercially important because Flash models are the ones developers can actually afford to call at scale. A breakthrough on an elite math problem is interesting for researchers, but a low-latency model that can carry more reasoning work is what changes product economics.
That is also why the comparison with Gemini 2.5 Flash and Gemini 3 Flash matters. Earlier Flash models drew attention for punching above their weight on math and coding, but the new chatter suggests Google has not eased off. It appears to be moving Flash upward while keeping Pro as the place for deeper, slower thinking. That split lets Google sell two different ideas at once: rapid response for everyday use and near-frontier reasoning for users who need more.
The reported GPT-5.5 Pro comparison adds another layer. OpenAI launched GPT-5.5 in April, positioning it as a stronger reasoning and professional work model, and the community chatter now places its Pro version in the same tiny group of models that can handle IMO 2025 Problem 6. If that survives wider testing, the race is no longer just about who can solve benchmark suites. It is about who can solve the hardest human-designed problems consistently, cleanly, and without a complex harness doing the real work.
Benchmarking gets harder
The problem with elite math benchmarks is that they are easy to fetishize and hard to interpret. A model can look extraordinary on one olympiad problem and still fail on adjacent tasks that require similar discipline. The more the industry leans on these demonstrations, the more pressure there will be to explain the setup, the prompting, the verification process, and whether a result came from the model itself or from surrounding tooling.
There is a deeper issue here for frontier labs. When a model solves a contest problem built to defeat strong human solvers, the marketing temptation is obvious. The research challenge is harder. Labs now need to show not just that a model can reach a stunning answer once, but that the capability is robust, reproducible, and not dependent on hidden scaffolding. That is where these results will be judged in the coming months.
For now, the signal is simple but still provisional. If community benchmarks are accurate, Gemini 3.2 Flash has joined an exceptionally small club, and it has done so from a model family meant to be fast. That would be a meaningful shift. It says the line between quick and advanced is getting thinner, and it puts even more weight on how Google, OpenAI, and everyone else choose to measure reasoning from here.
Also read: Gemini 2.5 Flash adds a new twist to the AI math race • AI's First Triage: Labor data show early contractions in high‑exposure white‑collar roles • OpenAI wins Musk lawsuit as jury rejects nonprofit betrayal claim
