Benchmarks · Coding & software engineering

Codeforces / CodeContests

also: Codeforces rating, CodeContests

Can a model solve genuinely novel algorithmic problems under contest conditions — the kind that require devising an approach, not recalling one — well enough to rank against real competitive programmers?

Codeforces (competitive-programming platform); CodeContests dataset curated by Google DeepMindReleased 2 February 2022Live

Codeforces is not an AI benchmark by design — it is a competitive-programming platform, running timed contests judged by strict hidden test cases, that has existed since long before large language models. It became one anyway, because its problems are genuinely novel each contest, its scoring is unforgiving, and its Elo-style rating system already gives a precise, human-calibrated scale to compare against. DeepMind’s AlphaCode made that comparison explicit in 2022, entering ten real Codeforces contests and ranking in roughly the top 54% of several-thousand-strong fields — DeepMind’s own description of the first AI system to reach a competitive standard in programming contests. Alongside the result, DeepMind released CodeContests, a curated dataset of some 13,000 problems drawn from Codeforces and similar judges, which became a standard training and evaluation resource in its own right.

What followed was one of the sharper capability curves in this record. OpenAI’s o1-preview reached roughly the 89th percentile in 2024, and o3 reported a 99.2nd-percentile rating that December — a jump from median-human to elite-human performance in under three years. As frontier reasoning models pushed further into territory with few human competitors to compare against, reporting shifted from contest percentile toward raw Elo figures quoted directly on the same numeric scale used for humans: Google reported Gemini 3 Deep Think reaching an Elo of 3455 in early 2026, within range of the strongest human grandmasters on the platform.

Because live Codeforces problems are set after any given model’s training cutoff, the platform has held up better against contamination concerns than most static coding benchmarks — though methodologies still vary: some labs enter live contests directly, others estimate a rating by scoring retrospectively against archived problems, and the two are not strictly comparable. It remains one of the few benchmarks in coding where a model’s score can be read on literally the same scale as a human’s.

The set

Codeforces is a long-running online platform hosting timed programming contests judged by strict correctness on hidden test cases; performance produces an Elo-style rating and, within a given contest, a percentile rank against several thousand human entrants. DeepMind's CodeContests, released alongside AlphaCode, packaged roughly 13,000 problems drawn from Codeforces and other judges (AtCoder, CodeChef, HackerEarth, Aizu) into a static training and evaluation set; live Codeforces contests remain the reference for headline 'rating' or 'percentile' claims because their problems postdate any fixed training cutoff.

Example

From CodeContests problem 1575A, 'Another Sorting Problem', verbatim: 'Andi and Budi... decided to combine their idea and sort it asc-desc-endingly, where the odd-indexed characters will be compared ascendingly, and the even-indexed characters will be compared descendingly.' One of thousands of such problems in the archive.huggingface.co

Where it stands

Reporting has shifted over time from contest percentile (AlphaCode's 54th percentile in 2022) to Elo-style ratings quoted directly against the human scale (Gemini 3 Deep Think's 3455 by early 2026, within reach of top human grandmasters); methodologies differ enough between labs — live contest submission versus retrospective scoring against archived problems — that ratings from different announcements are not strictly apples-to-apples.

How the top score changed hands

  1. February 2022AlphaCode~54th percentileAverage ranking across ten Codeforces contests with 5,000+ participants each — DeepMind's own description of the first AI system to reach a competitive standard in programming contests.
  2. September 2024OpenAI o1-preview~89th percentile
  3. December 2024OpenAI o399.2nd percentile
  4. February 2026Gemini 3 Deep Think (v2)Elo 3455

Current best: Gemini 3 Deep Think (v2) — Elo 3455 Google's own reported figure, alongside gold-medal-level results on the 2025 Math, Physics and Chemistry Olympiads; later 2026 models may score higher but a comparably sourced figure was not confirmed here.

In the timeline · 3 entries

More coding & software engineering benchmarks