Model benchmarks: capability and value.

The most capable model we price is GPT-6 Astra (index 169) at about $20.00 per million tokens. GPT-6.1 Terra comes within 6 points for $2.00, about 90% less.

Each dot is a model. The green dots are the efficient ones: nothing cheaper scores as high. Hover a dot for its numbers.

How little can you pay and still be close to the best?

Price is a typical chat mix: three parts input to one part output, in US dollars per million tokens, from the price list on this site. The comparison is this site's own working on Epoch's scores; a few points of index is a small difference, so use it as a guide, not a verdict.

Every model Epoch has scored, by when it was released. The green line follows the best score so far, and the labelled dots are the models that set a new record. Teal dots are open-weights models (anyone can download and run them); grey dots are closed.

Open versus closed. The best open-weights model, DeepSeek V4.1 Flash 1017, scores 159; the best closed one, GPT-6 Astra, scores 169: a gap of 10 points. A closed model (GPT-5.4 Pro) first reached the open model's level in March 2026, about 8 months before it.

Hover a dot for the model and its score. The index is Epoch's own scale, so read the shape of the line rather than any single number. Models Epoch labels neither open nor closed are shown faintly.

Each maker's best model in each area, on the 0 to 100 scale. Under each score is the model it belongs to and how many of Epoch's benchmarks in that area it has been tested on ("1 of 4" is thin evidence). Shading compares makers within a column only. Only makers with at least three scored models are shown.

"Not tested" means Epoch has not run any benchmark in that area on that maker's models, not that they scored zero. Areas are not comparable with one another, because each is built from different benchmarks.

Index is Epoch's Capabilities Index, with the range Epoch is confident in underneath. The area columns are this site's average of the model's Epoch results in that area, where 0 is what guessing would score and 100 is perfect; "2 of 4" means Epoch has run two of the four benchmarks in that area on the model, so compare areas with the same count with more confidence. Click a column heading to sort.