Each Epoch benchmark, plotted on the release date of every model that's been scored on it. As newer models arrive the ceiling of what's been achieved keeps rising — a real trend, distinct from and often mistaken for individual models improving over time.
One point per model at its release date. The line traces the best-known score at each moment.
Swipe sideways to see the full chart →
Model at its release dateRunning best score (frontier)Score confidence interval
Read this as a frontier chart, not a per-model trajectory. A fixed model's score on a fixed benchmark does not change over time — Claude 3.5 Sonnet's GPQA number today is the same one Anthropic reported at launch. The line rises because better models keep arriving, not because any specific model got better. Seeing one model's score change over time (as its evaluation is re-run with different scaffolds, for example) needs a different kind of data: daily snapshots of each board, which we will chart once several weeks of them have accumulated.
25 models have no known release date and are not plotted.
Benchmarks · frontier
Frontier milestones
Every time a ceiling moved: who, which benchmark, by how much. Tap a sparkline to pick a benchmark — the chart above follows it; tap an organisation to filter by lab. Each event is the diff of the running best the chart draws.